The present application relates generally to an improved data processing apparatus and method and more specifically to mechanisms for protecting cognitive systems from model stealing attacks.
Deep learning is part of a broader family of machine learning methods based on learning data representations as opposed to task-specific algorithms. Some representations are loosely based on interpretation of information processing and communication patterns in a biological nervous system, such as neural coding that attempts to define a relationship between various stimuli and associated neuronal responses in the brain. Research attempts to create efficient systems to learn these representations from large-scale, unlabeled data sets.
Deep learning architectures such as deep neural networks, deep belief networks, and recurrent neural networks, have been applied to fields including computer vision, speech recognition, natural language processing, audio recognition, social network filtering, machine translation, and bioinformatics where they produced results comparable to, and in some cases superior to, human experts.
Neural network based deep learning is a class of machine learning algorithms that use a cascade of many layers of nonlinear processing units for feature extraction and transformation. Each successive layer uses the output from the previous layer as input. The algorithms may be supervised or unsupervised and applications include pattern analysis (unsupervised) and classification (supervised). Neural network based deep learning is based on the learning of multiple levels of features or representations of the data with higher level features being derived from lower level features to form a hierarchical representation. The composition of a layer of nonlinear processing units of the neural network used in a deep learning algorithm depends on the problem to be solved. Layers that have been used in deep learning include hidden layers of an artificial neural network and sets of complicated propositional formulas. They may also include latent variables organized layer-wise in deep generative models such as the nodes in deep belief networks and deep Boltzmann machines.
This Summary is provided to introduce a selection of concepts in a simplified form that are further described herein in the Detailed Description. This Summary is not intended to identify key factors or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.
In one illustrative embodiment, a method for obfuscating training of trained cognitive model logic is provided, the method being performed in a data processing system comprising a processor and a memory, the memory comprising instructions executed by the processor to specifically configure the processor to implement a cognitive system comprising the trained cognitive model logic. The method comprises receiving, by the trained cognitive model logic of the cognitive system, input data for classification into one or more classes in a plurality of predefined classes as part of a cognitive operation of the cognitive system. The method also comprises processing, by the trained cognitive model logic, the input data by applying a trained cognitive model to the input data to generate an output vector having values for each of the plurality of predefined classes. Moreover, the method comprises modifying, by a perturbation insertion engine of the cognitive system, one or more values of the output vector by inserting a perturbation in a function associated with generating the output vector, to thereby generate a modified output vector. In addition, the method comprises outputting, by the trained cognitive model logic, the modified output vector. The perturbation modifies the one or more values to obfuscate the trained configuration of the trained cognitive model logic while maintaining accuracy of classification of the input data.
In other illustrative embodiments, a computer program product comprising a computer useable or readable medium having a computer readable program is provided. The computer readable program, when executed on a computing device, causes the computing device to perform various ones of, and combinations of, the operations outlined above with regard to the method illustrative embodiment.
In yet another illustrative embodiment, a system/apparatus is provided. The system/apparatus may comprise one or more processors and a memory coupled to the one or more processors. The memory may comprise instructions which, when executed by the one or more processors, cause the one or more processors to perform various ones of, and combinations of, the operations outlined above with regard to the method illustrative embodiment.
These and other features and advantages of the present invention will be described in, or will become apparent to those of ordinary skill in the art in view of, the following detailed description of the example embodiments of the present invention.
The invention, as well as a preferred mode of use and further obj ectives and advantages thereof, will best be understood by reference to the following detailed description of illustrative embodiments when read in conjunction with the accompanying drawings, wherein:
The illustrative embodiments provide mechanisms for protecting cognitive systems, such as those comprising neural networks, machine learning, and/or deep learning mechanisms, from model stealing attacks. While the illustrative embodiments will be described in the context of a neural network based mechanism and cognitive system, the illustrative embodiments are not limited to such. Rather, the mechanisms of the illustrative embodiments may be utilized with any artificial intelligence mechanism, machine learning mechanism, deep learning mechanism, or the like, whose output may be modified in accordance with the illustrative embodiments set forth hereafter to thereby obfuscate the training of the internal mechanisms, e.g., models, neural networks, or the like, such that the internal mechanisms are not able to be reproduced via a model stealing attack.
The illustrative embodiments introduce noise into the output of a protected cognitive system that prevents an external party from being able to reproduce the configuration and training of a cognitive system. That is, the noise obscures the actual output generated by the cognitive system while maintaining the correctness of the output. In this way, the cognitive system can be used to perform its operations while preventing others from generating their own version of the trained and configured cognitive system that would generate correct output. While an attacker may be able to assume the noisy output to be correct output for training their own cognitive system model, the attacker's model will still not generate the same output of the cognitive system model that the attacker is attempting to recreate. This is because even with a noisy output, the correct output may be one of multiple possibilities, and only the exact same cognitive system model architecture, with the same model weights, can correctly identify which of the multiple possibilities is the correct one.
The success of neural network based systems has resulted in many web services based on them. Service providers provide application program interfaces (APIs) to end users of the web services through which the end users may submit, via their client computing devices, input data to be processed by the web service, and are provided results data indicating the results of the operations of the web services on the input data. Many times, cognitive systems utilize the neural networks to perform classification type operations to classify input data into various defined categories of information. For example, in an image processing web service, an input image comprising a plurality of data points, e.g., pixels, may be input to the web service which operates on the input image data to classify elements of the input image into types of objects present within the image, e.g., the image comprises a person, a car, a building, a dog, etc., to thereby perform object or image recognition. Similar types of classification analysis may be performed for various other types of input data including, but not limited to, speech recognition, natural language processing, audio recognition, social network filtering, machine translation, and bioinformatics. Of particular interest to some illustrative embodiments described herein, such web services may provide functionality for analyzing patient information in patient electronic medical records (EMRs) using natural language processing, analyzing medical images such as x-ray images, magnetic resonance imaging (MRI) images, computed tomography (CT) scan images, etc., and the like.
In many cases, the service providers charge the end users a fee for the use of the web service that is provided by the implementation of a neural network based cognitive system. However, it has been recognized that end users may utilize the APIs provided by the service provider to submit sets of input data to acquire enough output data to replicate the training of the neural network of the cognitive system such that the end user is able to generate their own trained neural network and thereby avoid having to utilize the service provider's service, resulting in a loss of revenue for the service provider. That is, if the end user is utilizing the service provider's web service to label a set of input data based on a classification operation performed on the input data, after submitting enough input data sets, e.g., 10,000 input data sets (where the term “data set” as used herein refers to a set of one or more data samples), and obtaining the corresponding output labels, the output labels may be used as a “golden” set or ground truth for training another neural network, e.g., the end user's own neural network, to perform a similar classification operation. This is referred to herein as a model stealing attack in that an end user, referred to hereafter as an “attacker”, motivated to recreate the trained neural network surreptitiously, attempts to steal the neural network model created and trained by the service provider through utilization of the service provider's APIs.
The illustrative embodiments reduce or eliminate the attacker's ability to perform a model stealing attack by introducing perturbations, or noise, into the output probabilities generated by the neural network, so as to fool the attacker who is trying the copy the neural network model. The perturbations that are introduced deviate the attacker's gradients from a correct direction and amount and minimize loss in the accuracy of the protected neural network model. To satisfy these two criteria, in some illustrative embodiments, two general guidelines are followed when generating the perturbations: (1) using the perturbation, one or more ways of learning the machine learning parameters are added to cause ambiguity in the model stealing process, e.g., in one illustrative embodiment, the sign of the first order derivative is reversed (the first order derivative identifies the direction, e.g., increasing/decreasing, of a function or curve); and (2) noise is added primarily in either end of the function, e.g., in a softmax or sigmoid function, up to +/−0.5.
It should be appreciated that these are only example guidelines for some illustrative embodiments and many different modifications may be made to these without departing from the spirit and scope of the present invention. For example, the activation functions need not be softmax or sigmoid functions, as these were chosen for the illustrative embodiments due to their standard use for deep learning classifiers. Any activation function of the protected neural network may be utilized in which ambiguity is added by way of the introduction of noise in accordance with the illustrative embodiments.
In general, the illustrative embodiments provide various methods and mechanisms to add perturbations without negatively affecting the accuracy of the model or neural network. For example, any noise may be added to the activation function of the model or neural network to thereby cause ambiguity in the output that will fool a model stealing attacker. However, in some illustrative embodiments, the perturbation mechanisms of the illustrative embodiments add any noise that does not change the result classification, i.e. the output of the class with the highest probability, generate by the trained model or neural network. That is, given an output probability vector y=[y_1, . . . , y_n], the perturbation mechanisms of these illustrative embodiments may add noise d such that argmax_i {y_i+d_i}=argmax_i {y_i}.
For added ambiguity in the output of the trained model or neural network, the perturbation mechanisms of the illustrative embodiments may add noise that causes not only the ambiguity, but also a sign change of gradients. In such an embodiment, the perturbation mechanisms add larger perturbations to clear cases such that the probability of a class is close to either 1 or 0 and thus, the result class of an output is likely to be preserved. Also, by adding noise to these cases, the perturbation mechanisms add ambiguity of learning. Lastly, by the added noise, the direction of gradients are opposite to the original gradients as the original model has higher probability with clearer cases as oppose to the perturbed output having lower probability with clearer cases.
There may be many different implementations of the perturbations that satisfy this criteria and all such perturbations are considered to be within the spirit and scope of the present invention. That is any functions that generate perturbations in the output of the neural network which satisfy the above criteria and guidelines may be used without departing from the spirit and scope of the present invention.
For example, assume there is a given a neural network f(x)=sigma(h(x)) where sigma is a softmax or sigmoid function, h(x) is the function representing the remainder of the neural network, and x is the input data. Various possible perturbations satisfy the above criteria and guidelines, examples of which are as follows:
1. Copy-protect(f(x))=normalization(sigma(h(x))−0.5(sigma(0.25h(x)) 0.5));
2. Gaussian noise up to +/−0.5 on [h1,inf) and (−inf,−h1] where h1 is minimum h(x) such that sigma (h(x))>0.99; and
3. Random noise h(x) such that the ranking of dimensions of sigma(h(x)+h′(x)) is equal to sigma(h(x)).
where normalization is an identity function if sigma is the sigmoid function, or is a function that divides the input vector by the sum of its values if sigma is the softmax function.
In the example perturbation 1 above, ambiguous cases are kept as-is, however the perturbation makes the output of the model or neural network less certain if the output for the same input is more certain, up to 0.5 which keeps the result class the same. That is, the higher the probability/confidence for the original model, the lower the probability/confidence for the protected model.
In the example perturbation 2 above, a type of random noise (which is called Gaussian noise) is added when the output is certain or probable, i.e. the probability of the classification is high, e.g., 1.0, 0.9, etc. depending on the implementation. The difference between perturbation 1 above and perturbation 2 above is that perturbation 1 adds more noise if the result is more certain while perturbation 2 does not require such ordering and thus, the more certain outputs may have less noise in some cases. Perturbation 3 adds noise that does not change the relative order of likely classes given the input data, e.g., if the input image data is more likely to be a bird than a cow, for example, then the perturbation adds noise as long as this order is preserved.
These perturbations minimize the change in boundary case, such as f(x)=0.5, such that there is almost no change in this area, but if the probability score is high, e.g., closer to f(x)=1.0, or low, e.g., closer to f(x)=0.0, the perturbation is large. However, in this case, the ranking of the output classifications do not change because the change is up to +/−0.25 for the highest ranked classification (#1), and 1.0-0.25=0.75 is still the highest probability score among the classifications. That is, if the original probability of a classification is 1.0 and, through the introduction of noise in accordance with illustrative embodiments set forth herein, the probability is reduced to 0.75, then this may still be considered high and the classification of the input remains the same. However, if the probability were to drop to 0.5, then the average user would consider this to be uncertain, indicating that the output may not be useable and further analysis may be required.
Thus, the mechanisms of the illustrative embodiments improve the operation of the neural network and the cognitive system implementing the neural network, by adding additional non-generic functionality that previously did not exist in the neural network mechanism or cognitive system, specifically for avoiding model stealing attacks. The mechanisms of the illustrative embodiments add additional technological logic in the neural network and cognitive system that specifically implements the introduction of perturbations following the criteria and guidelines noted above to allow for obfuscation of the training of the neural network, machine learning model, deep learning model, or the like, while maintaining the usability of the resulting output, e.g., the classification and labeling of the output data is still accurate even though the actual probability values generated by the model are not accurate to the training of the model. The mechanisms of the illustrative embodiments are specific to a technological environment involving one or more data processing systems and/or computing devices that are specifically configured to implement the additional logic of the present invention thereby resulting in a non-generic technological environment comprising one or more non-generic data processing systems and/or computing devices. Moreover, the illustrative embodiments are specifically directed to solving the technological problem of model stealing attacks by reproducing the training of specialized computing devices having neural network models, machine learning models, deep learning models, or other such artificial intelligence or cognitive operation based computing mechanisms.
Before beginning the discussion of the various aspects of the illustrative embodiments in more detail, it should first be appreciated that throughout this description the term “mechanism” will be used to refer to elements of the present invention that perform various operations, functions, and the like. A “mechanism,” as the term is used herein, may be an implementation of the functions or aspects of the illustrative embodiments in the form of an apparatus, a procedure, or a computer program product. In the case of a procedure, the procedure is implemented by one or more devices, apparatus, computers, data processing systems, or the like. In the case of a computer program product, the logic represented by computer code or instructions embodied in or on the computer program product is executed by one or more hardware devices in order to implement the functionality or perform the operations associated with the specific “mechanism.” Thus, the mechanisms described herein may be implemented as specialized hardware, software executing on general purpose hardware, software instructions stored on a medium such that the instructions are readily executable by specialized or general purpose hardware, a procedure or method for executing the functions, or a combination of any of the above.
The present description and claims may make use of the terms “a”, “at least one of”, and “one or more of” with regard to particular features and elements of the illustrative embodiments. It should be appreciated that these terms and phrases are intended to state that there is at least one of the particular feature or element present in the particular illustrative embodiment, but that more than one can also be present. That is, these terms/phrases are not intended to limit the description or claims to a single feature/element being present or require that a plurality of such features/elements be present. To the contrary, these terms/phrases only require at least a single feature/element with the possibility of a plurality of such features/elements being within the scope of the description and claims.
Moreover, it should be appreciated that the use of the term “engine,” if used herein with regard to describing embodiments and features of the invention, is not intended to be limiting of any particular implementation for accomplishing and/or performing the actions, steps, processes, etc., attributable to and/or performed by the engine. An engine may be, but is not limited to, software, hardware and/or firmware or any combination thereof that performs the specified functions including, but not limited to, any use of a general and/or specialized processor in combination with appropriate software loaded or stored in a machine readable memory and executed by the processor. Further, any name associated with a particular engine is, unless otherwise specified, for purposes of convenience of reference and not intended to be limiting to a specific implementation. Additionally, any functionality attributed to an engine may be equally performed by multiple engines, incorporated into and/or combined with the functionality of another engine of the same or different type, or distributed across one or more engines of various configurations.
In addition, it should be appreciated that the following description uses a plurality of various examples for various elements of the illustrative embodiments to further illustrate example implementations of the illustrative embodiments and to aid in the understanding of the mechanisms of the illustrative embodiments. These examples intended to be non-limiting and are not exhaustive of the various possibilities for implementing the mechanisms of the illustrative embodiments. It will be apparent to those of ordinary skill in the art in view of the present description that there are many other alternative implementations for these various elements that may be utilized in addition to, or in replacement of, the examples provided herein without departing from the spirit and scope of the present invention.
The present invention may be a system, a method, and/or a computer program product. The computer program product may include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present invention.
The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium may be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.
Computer readable program instructions described herein can be downloaded to respective computing/processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and/or a wireless network. The network may comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and/or edge servers. A network adapter card or network interface in each computing/processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing/processing device.
Computer readable program instructions for carrying out operations of the present invention may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like, and conventional procedural programming languages, such as the “C” programming language or similar programming languages. The computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PLA) may execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present invention.
Aspects of the present invention are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer readable program instructions.
These computer readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and/or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function/act specified in the flowchart and/or block diagram block or blocks.
The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable apparatus, or other device implement the functions/acts specified in the flowchart and/or block diagram block or blocks.
The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts or carry out combinations of special purpose hardware and computer instructions.
As noted above, the present invention provides mechanisms for protecting cognitive systems, such as those comprising neural networks and/or deep learning mechanisms, from model stealing attacks.
The examples shown in
As shown in
This process may be repeated for a plurality of sets of input data 120 to generate a plurality of labeled data sets 140 (which may also comprise labels for one or more data samples). A labeled data set 140 is a set of output data generated by the trained neural network model 130 where the unlabeled input data is augmented with additional tags or labels of meaningful information for the particular cognitive operation for which the data is to be used. For example, in a patient treatment recommendation cognitive system, the labeled data may comprise labels, tags, or annotations that specify various medical concepts with which the data is associated, e.g., a disease, a treatment, a patient's age, a patient's gender, etc. In the depicted example, the operation of the neural network model 130 is to classify a portion of an input image specified in a set of input data 120 into one of 10 categories representing numerical values that the portion of the input image represents, e.g., classes “0” to “9”. Thus, the label that is affixed to a set of input data 120 may be a label of “O” or “1” or “2”, etc.
The attacker 110, having obtained a plurality of labeled data sets 140 based on a plurality of input data sets 120 may utilize this correspondence of inputs/outputs to train their own model 150 to replicated the trained neural network model 130. Once the attacker 110 has their own replicated version 150 of the trained neural network model 130, they no longer need to utilize the original trained neural network model 130 to obtain the labeled data set 140 for future input data sets 120 and can utilize their own replicated model 150. This causes the provider of the original trained neural network model 130 to lose revenue from fees that may be charged for the use of original trained neural network model 130. In addition, this may give rise to competitors to the service provider that surreptitiously obtain the benefit of the resource investment of the service provider with regard to training the neural network model 130, without actually having to make such a resource investment.
As shown in
In this depicted example, each slot of the vector output 135 corresponds to a possible classification from “0” to “9” indicating the possible numerical values that the portion of the input image may represent. The probability values may range from 0% (e.g., 0.0) to 100% (e.g., 1.0) and may have various levels of precision based on the particular implementation desired. Thus, if a label or classification of “1” has a probability value of 1.0 this indicates absolute confidence that the input data set 120 represents the numerical value of “1” and a probability of value of 0.0 indicates that the input data set 120 does not represent the corresponding value, i.e. the label of that vector slot does not apply to the input data set 120.
While this is a simple example used for illustrative purposes, it should be appreciated that the number of classifications and corresponding labels, as well as the corresponding vector output 135, may be quite complex. As another example, these classifications may be, for example, in a medical imaging application where internal structures of human anatomy are being classified in a patient's chest, e.g., an aorta, a heart valve, a left ventricle, right ventricle, lungs, etc. It should be appreciated that the vector output 135 may comprise any number of potential vector slots, or classifications, at various levels of granularity depending on the particular application and implementation, and the vector output 135 may be of various sizes correspondingly.
A highest probability value vector slot (or simply “slot”) in the vector output 135 may be selected to label the corresponding input data set 120. Thus, for example, assuming that the trained neural network model 130 is properly trained, an input data set 120 having an image of a numerical value of “2” will have an output vector 135 similar to that shown in
As shown in
The introduction of the perturbations, or noise, into the output of the trained neural network model 130 results in a modified or manipulated labeled data set 170 that is provided to the attacker 110 rather than the actual labeled data set 140 that would otherwise have been generated by the operation of the trained neural network model 130. If the attacker uses the manipulated labeled data set 170 to train the attackers own neural network model 150, the result will be an ill-replicated model with lower performance.
To illustrate the way in which the introduction of perturbations or noise into the output generated by the trained neural network model operates to obfuscate the training of the neural network model, consider the example diagrams in
The sigmoid or softmax function can be considered to be stretched or shrunk depending on the training of the neural network model by updating the model weights, but is predictable to an attacker 110 given a labeled data set 140, for example, through curve fitting or the like, i.e. the attacker attempts to learn the same curve used by the trained neural network model 130 based on the collection of input data sets 120 and corresponding output labeled data 140 obtained from the trained neural network model 130. Typically such learning of the curve requires the calculation of a gradient (e.g., change in y coordinate divided by change in x coordinate of a graphed curve) from points along the curve to know the direction and magnitude of the curvature of the curve.
With reference now to
Because of the nature of the softmax or sigmoid function curve, there is more area of the curve into which perturbations or noise may be added at the ends of the curve. Thus, the mechanisms of some illustrative embodiments utilize perturbation injection logic that introduce such perturbations in the ends of the curve near 0.0 and 1.0, i.e. very low and very high probability value areas as previously mentioned above, so as to make an attacker's attempt to utilize the output of the trained neural network model 130 to train their own neural network model result in a lower performance model. The introduction of the perturbations or noise at the ends of the curve can be facilitated by subtracting a sigmoid function or hyperbolic tangent function that has a higher absolute value in the ends of the curve, like 0.5(sigma(0.25h(x))−0.5)) in perturbation 1 mentioned above.
Thus, the illustrative embodiments provide mechanisms for obfuscating the training of a neural network, machine learning, deep learning, of other artificial intelligence/cognitive model by introducing noise into the output of such trained models in such a way as to maintain the accuracy of the output, yet manipulate the output values to make detection of the specific curve or function to which the model correlates difficult to reproduce. The introduction of the perturbations, or noise, is done so as to minimize the change in boundary cases but introduce large size perturbations in areas of the curve or function where the probability values are relatively high/low, e.g., near 1.0 and near 0.0 in the case of a sigmoid/softmax function. While such perturbations are introduced into these areas of the function or curve, the perturbations are sized such that that output classes do not change in the modified output because the perturbation modifications are limited to less than predetermined amount of change that will not modify the output classifications.
As noted above, the mechanisms of the illustrative embodiments are directed to protecting trained neural network models, machine learning models, deep learning models, and the like, implemented in specialized logic of specially configured computing devices, data processing systems, or the like, of a technological environment. As such, the illustrative embodiments may be utilized in many different types of data processing environments. In order to provide a context for the description of the specific elements and functionality of the illustrative embodiments,
It should be appreciated that classification of input data may result in a labeled set of data that has labels or annotations representing the corresponding classes into which the non-labeled input data set is classified. This may be an intermediate step in performing other cognitive operations by the cognitive system that support decision making by human users, e.g., the cognitive system may be a decision support system. For example, in a medical domain, the cognitive system may operate to perform medical image analysis to identify anomalies for identification to a clinician, patient diagnosis and/or treatment recommendation, drug interaction analysis, or any of a plethora of other possible decision support operations.
It should be appreciated that the cognitive system, while shown as having a single request processing pipeline in the examples hereafter, may in fact have multiple request processing pipelines. Each request processing pipeline may be separately trained and/or configured to process requests associated with different domains or be configured to perform the same or different analysis on input requests (or questions in implementations using a QA pipeline), depending on the desired implementation. For example, in some cases, a first request processing pipeline may be trained to operate on input requests directed to a medical image analysis, while a second request processing pipeline may be configured and trained to operate on input requests concerning patient electronic medical record (EMR) analysis involving natural language processing. In other cases, for example, the request processing pipelines may be configured to provide different types of cognitive functions or support different types of applications, such as one request processing pipeline being used for patient treatment recommendation generation, while another pipeline may be trained for financial industry based forecasting, etc.
Moreover, each request processing pipeline may have their own associated corpus or corpora that they ingest and operate on, e.g., one corpus for medical treatment documents and another corpus for financial industry domain related documents in the above examples. In some cases, the request processing pipelines may each operate on the same domain of input questions but may have different configurations, e.g., different annotators or differently trained annotators, such that different analysis and potential answers are generated. The cognitive system may provide additional logic for routing input questions to the appropriate request processing pipeline, such as based on a determined domain of the input request, combining and evaluating final results generated by the processing performed by multiple request processing pipelines, and other control and interaction logic that facilitates the utilization of multiple request processing pipelines.
As noted above, one type of request processing pipeline with which the mechanisms of the illustrative embodiments may be utilized is a Question Answering (QA) pipeline. The description of example embodiments of the present invention hereafter will utilize a QA pipeline as an example of a request processing pipeline that may be augmented to include mechanisms in accordance with one or more illustrative embodiments. It should be appreciated that while the present invention will be described in the context of the cognitive system implementing one or more QA pipelines that operate on an input question, the illustrative embodiments are not limited to such. Rather, the mechanisms of the illustrative embodiments may operate on requests that are not posed as “questions” but are formatted as requests for the cognitive system to perform cognitive operations on a specified set of input data using the associated corpus or corpora and the specific configuration information used to configure the cognitive system. For example, rather than asking a natural language question of “What diagnosis applies to patient P?”, the cognitive system may instead receive a request of “generate diagnosis for patient P,” or the like. It should be appreciated that the mechanisms of the QA system pipeline may operate on requests in a similar manner to that of input natural language questions with minor modifications. In fact, in some cases, a request may be converted to a natural language question for processing by the QA system pipelines if desired for the particular implementation.
As will be discussed in greater detail hereafter, the illustrative embodiments may be integrated in, augment, and extend the functionality of these QA pipeline, or request processing pipeline, mechanisms to protect the models implemented in these pipelines, or by the cognitive system as a whole, from model stealing attacks. In particular, in portions of the cognitive system in which the trained neural network models, machine learning models, deep learning models, or the like, are employed to generate labeled data set outputs, the mechanisms of the illustrative embodiments may be implemented to modify the labeled data set outputs by the introduction of noise into the probability values generated by the trained models and thereby obfuscate the training of the models.
As the mechanisms of the illustrative embodiments may be part of a cognitive system and may improve the operation of the cognitive system by protecting it from model stealing attacks, it is important to first have an understanding of how cognitive systems and question and answer creation in a cognitive system implementing a QA pipeline is implemented before describing how the mechanisms of the illustrative embodiments are integrated in and augment such cognitive systems and request processing pipeline, or QA pipeline, mechanisms. It should be appreciated that the mechanisms described in
As an overview, a cognitive system is a specialized computer system, or set of computer systems, configured with hardware and/or software logic (in combination with hardware logic upon which the software executes) to emulate human cognitive functions. These cognitive systems apply human-like characteristics to conveying and manipulating ideas which, when combined with the inherent strengths of digital computing, can solve problems with high accuracy and resilience on a large scale. A cognitive system performs one or more computer-implemented cognitive operations that approximate a human thought process as well as enable people and machines to interact in a more natural manner so as to extend and magnify human expertise and cognition. A cognitive system comprises artificial intelligence logic, such as natural language processing (NLP) based logic, for example, and machine learning logic, which may be provided as specialized hardware, software executed on hardware, or any combination of specialized hardware and software executed on hardware. This logic may implement one or more models, such as a neural network model, a machine learning model, a deep learning model, that may be trained for particular purposes for supporting the particular cognitive operations performed by the cognitive system. In accordance with the mechanisms of the illustrative embodiments, the logic further implements the perturbation insertion engine mechanisms described above, and hereafter, for introducing perturbations, or noise, into the outputs of the implemented models so as to obfuscate the training of the models to those who would attempt to perform model stealing attacks.
The logic of the cognitive system implements the cognitive computing operation(s), examples of which include, but are not limited to, question answering, identification of related concepts within different portions of content in a corpus, intelligent search algorithms, such as Internet web page searches, for example, medical diagnostic and treatment recommendations, other types of recommendation generation, e.g., items of interest to a particular user, potential new contact recommendations, etc., image analysis, audio analysis, and the like. The types and number of cognitive operations that may be implemented using the cognitive system of the illustrative embodiments are vast and cannot all be documented herein. Any cognitive computing operation emulating decision making and analysis performed by human beings, but in an artificial intelligence or cognitive computing manner, is intended to be within the spirit and scope of the present invention.
IBM Watson™ is an example of one such cognitive computing system which can process human readable language and identify inferences between text passages with human-like high accuracy at speeds far faster than human beings and on a larger scale. In general, such cognitive systems are able to perform the following functions:
In one aspect, cognitive computing systems (or simply “cognitive systems”) provide mechanisms for answering questions posed to these cognitive systems using a Question Answering pipeline or system (QA system) and/or process requests which may or may not be posed as natural language questions. The QA pipeline or system is an artificial intelligence application executing on data processing hardware that answers questions pertaining to a given subject-matter domain presented in natural language. The QA pipeline receives inputs from various sources including input over a network, a corpus of electronic documents or other data, data from a content creator, information from one or more content users, and other such inputs from other possible sources of input. Data storage devices store the corpus of data. A content creator creates content in a document for use as part of a corpus of data with the QA pipeline. The document may include any file, text, article, or source of data for use in the QA system. For example, a QA pipeline accesses a body of knowledge about the domain, or subject matter area, e.g., financial domain, medical domain, legal domain, etc., where the body of knowledge (knowledgebase) can be organized in a variety of configurations, e.g., a structured repository of domain-specific information, such as ontologies, or unstructured data related to the domain, or a collection of natural language documents about the domain.
Content users input questions to cognitive system which implements the QA pipeline. The QA pipeline then answers the input questions using the content in the corpus of data by evaluating documents, sections of documents, portions of data in the corpus, or the like. When a process evaluates a given section of a document for semantic content, the process can use a variety of conventions to query such document from the QA pipeline, e.g., sending the query to the QA pipeline as a well-formed question which is then interpreted by the QA pipeline and a response is provided containing one or more answers to the question. Semantic content is content based on the relation between signifiers, such as words, phrases, signs, and symbols, and what they stand for, their denotation, or connotation. In other words, semantic content is content that interprets an expression, such as by using Natural Language Processing.
As will be described in greater detail hereafter, the QA pipeline receives an input question, parses the question to extract the major features of the question, uses the extracted features to formulate queries, and then applies those queries to the corpus of data. Based on the application of the queries to the corpus of data, the QA pipeline generates a set of hypotheses, or candidate answers to the input question, by looking across the corpus of data for portions of the corpus of data that have some potential for containing a valuable response to the input question. The QA pipeline then performs deep analysis on the language of the input question and the language used in each of the portions of the corpus of data found during the application of the queries using a variety of reasoning algorithms. There may be hundreds or even thousands of reasoning algorithms applied, each of which performs different analysis, e.g., comparisons, natural language analysis, lexical analysis, or the like, and generates a score. For example, some reasoning algorithms may look at the matching of terms and synonyms within the language of the input question and the found portions of the corpus of data. Other reasoning algorithms may look at temporal or spatial features in the language, while others may evaluate the source of the portion of the corpus of data and evaluate its veracity.
The scores obtained from the various reasoning algorithms indicate the extent to which the potential response is inferred by the input question based on the specific area of focus of that reasoning algorithm. Each resulting score is then weighted against a statistical model. The statistical model captures how well the reasoning algorithm performed at establishing the inference between two similar passages for a particular domain during the training period of the QA pipeline. The statistical model is used to summarize a level of confidence that the QA pipeline has regarding the evidence that the potential response, i.e. candidate answer, is inferred by the question. This process is repeated for each of the candidate answers until the QA pipeline identifies candidate answers that surface as being significantly stronger than others and thus, generates a final answer, or ranked set of answers, for the input question.
As mentioned above, QA pipeline mechanisms operate by accessing information from a corpus of data or information (also referred to as a corpus of content), analyzing it, and then generating answer results based on the analysis of this data. Accessing information from a corpus of data typically includes: a database query that answers questions about what is in a collection of structured records, and a search that delivers a collection of document links in response to a query against a collection of unstructured data (text, markup language, etc.). Conventional question answering systems are capable of generating answers based on the corpus of data and the input question, verifying answers to a collection of questions for the corpus of data, correcting errors in digital text using a corpus of data, and selecting answers to questions from a pool of potential answers, i.e. candidate answers.
Content creators, such as article authors, electronic document creators, web page authors, document database creators, and the like, determine use cases for products, solutions, and services described in such content before writing their content. Consequently, the content creators know what questions the content is intended to answer in a particular topic addressed by the content. Categorizing the questions, such as in terms of roles, type of information, tasks, or the like, associated with the question, in each document of a corpus of data allows the QA pipeline to more quickly and efficiently identify documents containing content related to a specific query. The content may also answer other questions that the content creator did not contemplate that may be useful to content users. The questions and answers may be verified by the content creator to be contained in the content for a given document. These capabilities contribute to improved accuracy, system performance, machine learning, and confidence of the QA pipeline. Content creators, automated tools, or the like, annotate or otherwise generate metadata for providing information useable by the QA pipeline to identify these question and answer attributes of the content.
Operating on such content, the QA pipeline generates answers for input questions using a plurality of intensive analysis mechanisms which evaluate the content to identify the most probable answers, i.e. candidate answers, for the input question. The most probable answers are output as a ranked listing of candidate answers ranked according to their relative scores or confidence measures calculated during evaluation of the candidate answers, as a single final answer having a highest ranking score or confidence measure, or which is a best match to the input question, or a combination of ranked listing and final answer.
The cognitive system 300 is configured to implement a request processing pipeline 308 that receive inputs from various sources. The requests may be posed in the form of a natural language question, natural language request for information, natural language request for the performance of a cognitive operation, or the like. For example, the cognitive system 300 receives input from the network 302, a corpus or corpora of electronic documents 306, cognitive system users, and/or other data and other possible sources of input. In one embodiment, some or all of the inputs to the cognitive system 300 are routed through the network 302. The various computing devices 304A-D on the network 302 include access points for content creators and cognitive system users. Some of the computing devices 304A-D include devices for a database storing the corpus or corpora of data 306 (which is shown as a separate entity in
In one embodiment, the content creator creates content in a document of the corpus or corpora of data 306 for use as part of a corpus of data with the cognitive system 300. The document includes any file, text, article, or source of data for use in the cognitive system 300. Cognitive system users access the cognitive system 300 via a network connection or an Internet connection to the network 302, and input questions/requests to the cognitive system 300 that are answered/processed based on the content in the corpus or corpora of data 306. In one embodiment, the questions/requests are formed using natural language. The cognitive system 300 parses and interprets the question/request via a pipeline 308, and provides a response to the cognitive system user, e.g., cognitive system user 310, containing one or more answers to the question posed, response to the request, results of processing the request, or the like. In some embodiments, the cognitive system 300 provides a response to users in a ranked list of candidate answers/responses while in other illustrative embodiments, the cognitive system 300 provides a single final answer/response or a combination of a final answer/response and ranked listing of other candidate answers/responses.
The cognitive system 300 implements the pipeline 308 which comprises a plurality of stages for processing an input question/request based on information obtained from the corpus or corpora of data 306. The pipeline 308 generates answers/responses for the input question or request based on the processing of the input question/request and the corpus or corpora of data 306. The pipeline 308 will be described in greater detail hereafter with regard to
In some illustrative embodiments, the cognitive system 300 may be the IBM Watson™ cognitive system available from International Business Machines Corporation of Armonk, N.Y., which is augmented with the mechanisms of the illustrative embodiments described hereafter. As outlined previously, a pipeline of the IBM Watson™ cognitive system receives an input question or request which it then parses to extract the maj or features of the question/request, which in turn are then used to formulate queries that are applied to the corpus or corpora of data 306. Based on the application of the queries to the corpus or corpora of data 306, a set of hypotheses, or candidate answers/responses to the input question/request, are generated by looking across the corpus or corpora of data 306 for portions of the corpus or corpora of data 306 (hereafter referred to simply as the corpus 306) that have some potential for containing a valuable response to the input question/response (hereafter assumed to be an input question). The pipeline 308 of the IBM Watson™ cognitive system then performs deep analysis on the language of the input question and the language used in each of the portions of the corpus 306 found during the application of the queries using a variety of reasoning algorithms.
The scores obtained from the various reasoning algorithms are then weighted against a statistical model that summarizes a level of confidence that the pipeline 308 of the IBM Watson™ cognitive system 300, in this example, has regarding the evidence that the potential candidate answer is inferred by the question. This process is be repeated for each of the candidate answers to generate ranked listing of candidate answers which may then be presented to the user that submitted the input question, e.g., a user of client computing device 310, or from which a final answer is selected and presented to the user. More information about the pipeline 308 of the IBM Watson™ cognitive system 300 may be obtained, for example, from the IBM Corporation website, IBM Redbooks, and the like. For example, information about the pipeline of the IBM Watson™ cognitive system can be found in Yuan et al., “Watson and Healthcare,” IBM developerWorks, 2011 and “The Era of Cognitive Systems: An Inside Look at IBM Watson and How it Works” by Rob High, IBM Redbooks, 2012.
As noted above, while the input to the cognitive system 300 from a client device may be posed in the form of a natural language question, the illustrative embodiments are not limited to such. Rather, the input question may in fact be formatted or structured as any suitable type of request which may be parsed and analyzed using structured and/or unstructured input analysis, including but not limited to the natural language parsing and analysis mechanisms of a cognitive system such as IBM Watson™, to determine the basis upon which to perform cognitive analysis and providing a result of the cognitive analysis.
Regardless of the manner by which the question or request is input to the cognitive system 300, the processing of the request or question involves the application of a trained model, e.g., neural network model, machine learning model, deep learning model, etc., to an input data set as described previously above. This input data set may represent features of the actual request or question itself, data submitted along with the request or question upon which processing is to be performed, or the like. The application of the trained model to an input data set may occur at various points during the performance of the cognitive computing operations by the cognitive system. For example, the trained model may be utilized during feature extraction and classification by a feature extraction stage of processing of the request or input question, e.g., taking a natural language term in the request or question and classifying it as one of a plurality of possible concepts that the term corresponds to, e.g., classifying the term “truck” in an input question or request into a plurality of possible classes, one of which may be “vehicle”. As another example, a portion of an image comprising a plurality of pixel data may have the trained model applied to it to determine what the object is that is in the portion of the image. The mechanisms of the illustrative embodiments operate on the output of the trained model as discussed previously above, which may be an intermediate operation within the overall cognitive system's cognitive computing operations, e.g., classification of a portion of a medical image into one of a plurality of different anatomical structures may be an intermediate operation to performing an anomaly identification and treatment recommendation cognitive computing operation.
As shown in
In one illustrative embodiment, the perturbation insertion engine 320 meets these criteria by using a perturbation function that reverses the sign of the first order derivative of the output probability curve, e.g., the sigmoid or softmax curve of the probability values, and adds noise or perturbations at the ends of the curve, near the maximum and minimum values of the curve, up to +/−half the range from the minimum value to the maximum value, e.g., in the case of a softmax or sigmoid probability value curve that ranges from 0% to 100%, the noise or perturbations have a magnitude of up to +/−0.5. As mentioned previously above, the particular perturbation function utilized may take many different forms including those previously listed above and others that meet the criteria and guidelines mentioned above.
The resulting modified output vector provides modified probability values while maintaining the correctness of the classification and associated labels that are associated with the input data in a labeled data set. Thus, correct classification and labeling of the input data set is still performed while obfuscating the actual training of the trained model logic 360. The resulting classified or labeled data set may be provided to further stages of processing downstream in the pipeline 306 for further processing and performance of the overall cognitive operation for which the cognitive system 300 is employed.
Thus, an attacker, such as a user of client computing device 310 or the like, is not able to submit a plurality of input data sets, obtain corresponding labeled output data sets and corresponding probability values of output vectors, and thereby train their own trained models to accurately replicate the training of the trained model logic 360 by utilizing the labeled data set and its associated probability values in the vector output as training data. To the contrary, doing so would result in a model that provides significantly lower performance than that of the trained model logic 360 resulting in a need to continue to utilize the trained model logic 360. In the case where a service provider charges a fee for utilization of the cognitive system 300 and/or trained model logic 360, this will result in a continued revenue stream for the service provider.
It should be appreciated that while
As noted above, the mechanisms of the illustrative embodiments are rooted in the computer technology arts and are implemented using logic present in such computing or data processing systems. These computing or data processing systems are specifically configured, either through hardware, software, or a combination of hardware and software, to implement the various operations described above. As such,
In the depicted example, data processing system 400 employs a hub architecture including north bridge and memory controller hub (NB/MCH) 402 and south bridge and input/output (I/O) controller hub (SB/ICH) 404. Processing unit 406, main memory 408, and graphics processor 410 are connected to NB/MCH 402. Graphics processor 410 is connected to NB/MCH 402 through an accelerated graphics port (AGP).
In the depicted example, local area network (LAN) adapter 412 connects to SB/ICH 404. Audio adapter 416, keyboard and mouse adapter 420, modem 422, read only memory (ROM) 424, hard disk drive (HDD) 426, CD-ROM drive 430, universal serial bus (USB) ports and other communication ports 432, and PCI/PCIe devices 434 connect to SB/ICH 404 through bus 438 and bus 440. PCI/PCIe devices may include, for example, Ethernet adapters, add-in cards, and PC cards for notebook computers. PCI uses a card bus controller, while PCIe does not. ROM 424 may be, for example, a flash basic input/output system (BIOS).
HDD 426 and CD-ROM drive 430 connect to SB/ICH 404 through bus 440. HDD 426 and CD-ROM drive 430 may use, for example, an integrated drive electronics (IDE) or serial advanced technology attachment (SATA) interface. Super I/O (SIO) device 436 is connected to SB/ICH 404.
An operating system runs on processing unit 406. The operating system coordinates and provides control of various components within the data processing system 400 in
As a server, data processing system 400 may be, for example, an IBM® eServer™ System p® computer system, running the Advanced Interactive Executive (AIX®) operating system or the LINUX® operating system. Data processing system 400 may be a symmetric multiprocessor (SMP) system including a plurality of processors in processing unit 406. Alternatively, a single processor system may be employed.
Instructions for the operating system, the object-oriented programming system, and applications or programs are located on storage devices, such as HDD 426, and are loaded into main memory 408 for execution by processing unit 406. The processes for illustrative embodiments of the present invention are performed by processing unit 406 using computer usable program code, which is located in a memory such as, for example, main memory 408, ROM 424, or in one or more peripheral devices 426 and 430, for example.
A bus system, such as bus 438 or bus 440 as shown in
Those of ordinary skill in the art will appreciate that the hardware depicted in
Moreover, the data processing system 400 may take the form of any of a number of different data processing systems including client computing devices, server computing devices, a tablet computer, laptop computer, telephone or other communication device, a personal digital assistant (PDA), or the like. In some illustrative examples, data processing system 400 may be a portable computing device that is configured with flash memory to provide non-volatile memory for storing operating system files and/or user-generated data, for example. Essentially, data processing system 400 may be any known or later developed data processing system without architectural limitation.
The QA system pipeline of
As shown in
In addition, the extracted major features include key words and phrases classified into question characteristics, such as the focus of the question, the lexical answer type (LAT) of the question, and the like. As referred to herein, a lexical answer type (LAT) is a word in, or a word inferred from, the input question that indicates the type of the answer, independent of assigning semantics to that word. For example, in the question “What maneuver was invented in the 1500s to speed up the game and involves two pieces of the same color?,” the LAT is the string “maneuver.” The focus of a question is the part of the question that, if replaced by the answer, makes the question a standalone statement. For example, in the question “What drug has been shown to relieve the symptoms of ADD with relatively few side effects?,” the focus is “drug” since if this word were replaced with the answer, e.g., the answer “Adderall” can be used to replace the term “drug” to generate the sentence “Adderall has been shown to relieve the symptoms of ADD with relatively few side effects.” The focus often, but not always, contains the LAT. On the other hand, in many cases it is not possible to infer a meaningful LAT from the focus.
The classification of the extracted features from the input question may be performed using one or more trained models 525 which may be implemented, for example, as neural network models, machine learning models, deep learning models, or other type of artificial intelligence based model. As noted above, the mechanisms of the illustrative embodiments may be implemented at the question and topic analysis stage 520 with regard to the classification of the extracted features of the input question by such trained models 525. That is, as the trained model 525 operates on the input data, e.g., the extracted features from the input question, to classify the input data, prior to output of the vector output, the perturbation insertion engine 590 of the illustrative embodiments may operate to introduce perturbations into the probability values generated in the output vector while maintaining the accuracy of the classification as discussed above. Thus, while correct classification is still provided downstream along the QA system pipeline 500, any attacker obtaining access to the output vector probability values for purposes of training their own model using a model stealing attack, would be presented with inaccurate probability values that result in any model trained on such probability values providing lower performance than the trained model 525.
It should be appreciated that the input data, in some illustrative embodiments, need not be a formulated request or question, either structure or unstructured, but instead may simply be an input data set that is input with the implied request that the input data set be processed by the pipeline 500. For example, in embodiments where the pipeline 500 is configured to perform image analysis cognitive operations, input images may be provided as input to the pipeline 500 which extracts major features of the input images, classifies them according to the trained model 525, and performs other processing of the pipeline 500 as described hereafter to score the hypotheses as to what is shown in the image, to thereby generate a final result output. In other cases, audio input data may also be analyzed in a similar manner. Regardless of the nature of the input data being processed, the mechanisms of the illustrative embodiments may be employed to insert perturbations in the probability values associated with the classification operations performed by the trained models 525 so as to obfuscate the training of the trained models.
Referring again to
The queries are applied to one or more databases storing information about the electronic texts, documents, articles, websites, and the like, that make up the corpus of data/information, e.g., the corpus of data 306 in
The QA pipeline 500, in stage 550, then performs a deep analysis and comparison of the language of the input question and the language of each hypothesis or “candidate answer,” as well as performs evidence scoring to evaluate the likelihood that the particular hypothesis is a correct answer for the input question. As mentioned above, this involves using a plurality of reasoning algorithms, each performing a separate type of analysis of the language of the input question and/or content of the corpus that provides evidence in support of, or not in support of, the hypothesis. Each reasoning algorithm generates a score based on the analysis it performs which indicates a measure of relevance of the individual portions of the corpus of data/information extracted by application of the queries as well as a measure of the correctness of the corresponding hypothesis, i.e. a measure of confidence in the hypothesis. There are various ways of generating such scores depending upon the particular analysis being performed. In generally, however, these algorithms look for particular terms, phrases, or patterns of text that are indicative of terms, phrases, or patterns of interest and determine a degree of matching with higher degrees of matching being given relatively higher scores than lower degrees of matching.
Thus, for example, an algorithm may be configured to look for the exact term from an input question or synonyms to that term in the input question, e.g., the exact term or synonyms for the term “movie,” and generate a score based on a frequency of use of these exact terms or synonyms. In such a case, exact matches will be given the highest scores, while synonyms may be given lower scores based on a relative ranking of the synonyms as may be specified by a subject matter expert (person with knowledge of the particular domain and terminology used) or automatically determined from frequency of use of the synonym in the corpus corresponding to the domain. Thus, for example, an exact match of the term “movie” in content of the corpus (also referred to as evidence, or evidence passages) is given a highest score. A synonym of movie, such as “motion picture” may be given a lower score but still higher than a synonym of the type “film” or “moving picture show.” Instances of the exact matches and synonyms for each evidence passage may be compiled and used in a quantitative function to generate a score for the degree of matching of the evidence passage to the input question.
Thus, for example, a hypothesis or candidate answer to the input question of “What was the first movie?” is “The Horse in Motion.” If the evidence passage contains the statements “The first motion picture ever made was ‘The Horse in Motion’ in 1878 by Eadweard Muybridge. It was a movie of a horse running,” and the algorithm is looking for exact matches or synonyms to the focus of the input question, i.e. “movie,” then an exact match of “movie” is found in the second sentence of the evidence passage and a highly scored synonym to “movie,” i.e. “motion picture,” is found in the first sentence of the evidence passage. This may be combined with further analysis of the evidence passage to identify that the text of the candidate answer is present in the evidence passage as well, i.e. “The Horse in Motion.” These factors may be combined to give this evidence passage a relatively high score as supporting evidence for the candidate answer “The Horse in Motion” being a correct answer.
It should be appreciated that this is just one simple example of how scoring can be performed. Many other algorithms of various complexity may be used to generate scores for candidate answers and evidence without departing from the spirit and scope of the present invention.
In the synthesis stage 560, the large number of scores generated by the various reasoning algorithms are synthesized into confidence scores or confidence measures for the various hypotheses. This process involves applying weights to the various scores, where the weights have been determined through training of the statistical model employed by the QA pipeline 500 and/or dynamically updated. For example, the weights for scores generated by algorithms that identify exactly matching terms and synonym may be set relatively higher than other algorithms that are evaluating publication dates for evidence passages. The weights themselves may be specified by subject matter experts or learned through machine learning processes that evaluate the significance of characteristics evidence passages and their relative importance to overall candidate answer generation.
The weighted scores are processed in accordance with a statistical model generated through training of the QA pipeline 500 that identifies a manner by which these scores may be combined to generate a confidence score or measure for the individual hypotheses or candidate answers. This confidence score or measure summarizes the level of confidence that the QA pipeline 500 has about the evidence that the candidate answer is inferred by the input question, i.e. that the candidate answer is the correct answer for the input question.
The resulting confidence scores or measures are processed by a final confidence merging and ranking stage 570 which compares the confidence scores and measures to each other, compares them against predetermined thresholds, or performs any other analysis on the confidence scores to determine which hypotheses/candidate answers are the most likely to be the correct answer to the input question. The hypotheses/candidate answers are ranked according to these comparisons to generate a ranked listing of hypotheses/candidate answers (hereafter simply referred to as “candidate answers”). From the ranked listing of candidate answers, at stage 580, a final answer and confidence score, or final set of candidate answers and confidence scores, are generated and output to the submitter of the original input question via a graphical user interface or other mechanism for outputting information.
Thus, the illustrative embodiments provide mechanisms for protecting trained artificial intelligence or cognitive models, such as neural network model, from model stealing attacks. The illustrative embodiments introduce perturbations, or noise, in the probability values output by the trained models such that an attacker's calculation of gradients based on the output probability values is deviated from the correct direction and magnitude while minimizing loss in the accuracy of the trained model's classification or labeled data set. In some illustrative embodiments, this result is achieved by using a perturbation function that reverses the sign of the first order derivative of the sigmoid or softmax function of the trained model and adds noise or perturbations at the ends of the curve of the sigmoid or softmax function, near the minimum and maximum values of the curve. The result is that if an attacker uses the modified probability values output by the trained model as a basis for training their own model, the resulting attacker model will have less accuracy than the trained model it is attempting to replicate.
As noted above, it should be appreciated that the illustrative embodiments may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment containing both hardware and software elements. In one example embodiment, the mechanisms of the illustrative embodiments are implemented in software or program code, which includes but is not limited to firmware, resident software, microcode, etc.
A data processing system suitable for storing and/or executing program code will include at least one processor coupled directly or indirectly to memory elements through a communication bus, such as a system bus, for example. The memory elements can include local memory employed during actual execution of the program code, bulk storage, and cache memories which provide temporary storage of at least some program code in order to reduce the number of times code must be retrieved from bulk storage during execution. The memory may be of various types including, but not limited to, ROM, PROM, EPROM, EEPROM, DRAM, SRAM, Flash memory, solid state memory, and the like.
Input/output or I/O devices (including but not limited to keyboards, displays, pointing devices, etc.) can be coupled to the system either directly or through intervening wired or wireless I/O interfaces and/or controllers, or the like. I/O devices may take many different forms other than conventional keyboards, displays, pointing devices, and the like, such as for example communication devices coupled through wired or wireless connections including, but not limited to, smart phones, tablet computers, touch screen devices, voice recognition devices, and the like. Any known or later developed I/O device is intended to be within the scope of the illustrative embodiments.
Network adapters may also be coupled to the system to enable the data processing system to become coupled to other data processing systems or remote printers or storage devices through intervening private or public networks. Modems, cable modems and Ethernet cards are just a few of the currently available types of network adapters for wired communications. Wireless communication based network adapters may also be utilized including, but not limited to, 802.11 a/b/g/n wireless communication adapters, Bluetooth wireless adapters, and the like. Any known or later developed network adapters are intended to be within the spirit and scope of the present invention.
The description of the present invention has been presented for purposes of illustration and description, and is not intended to be exhaustive or limited to the invention in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The embodiment was chosen and described in order to best explain the principles of the invention, the practical application, and to enable others of ordinary skill in the art to understand the invention for various embodiments with various modifications as are suited to the particular use contemplated. The terminology used herein was chosen to best explain the principles of the embodiments, the practical application or technical improvement over technologies found in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.