Patent Application
20210303675
The subject matter described herein is directed to mitigation the effects of adversarial techniques such as score fuzzing and string stuffing which can act to cause machine learning model to misclassify a particular artifact.
Machine learning and other types of artificial intelligence models are being increasingly deployed across different applications and industries. Such models provide classifications which can be based, for example, on historical data with known outcomes or features. The classifications provided by such models (i.e., the model outputs, etc.) can take various forms including a Boolean output (e.g., good/bad, etc.), a numerical score (e.g., 0.00 to 1, 1 to 100, etc.), or a grouping (e.g., automobile, pedestrian, crosswalk, etc.). With some software implementations, the outputs of such models can be intercepted even when part of a larger workflow. Such interception can allow a malicious actor to manipulate the classification by such models by repeatedly modifying sample input data until a desired classification is received (even if such classification is not ultimately accurate).
In a first aspect, an artifact is received from which features are extracted so as to populate a vector. The features in the vector can be reduced using a feature reduction operations to result in a modified vector having a plurality of buckets. A presence of predetermined types of features are identified within buckets of the modified vector influencing a score above a pre-determined threshold. A contribution of the identified features within the high influence buckets of the modified vector is then attenuated. The modified vector is input into a classification model to generate a score which can be provided to a consuming application or process.
The classification model can characterize the artifact as being malicious or benign to access, execute, or continue to execute. In such cases, access or execution of the artifact can be prevented, reduced, and the like when the classification model characterizes the artifact as being malicious.
The classification model can be a machine learning model trained using a training data set and providing a continuous scale output. The machine learning model can take various forms including one or more a logistic regression model, a neural network, a convolutional neural network, a recurrent neural network, a generative adversarial network, a support vector machine, a random forest, or a Bayesian model.
The predetermined types of features can take different forms including alphanumeric strings.
A plurality of vectorized malware samples can be inputted into the classification model after which a plurality of scores based on the inputted vectorized malware samples are obtained. Based on these classification, the predetermined types of features within buckets of the vectorized malware samples that influence the scores above the pre-determined threshold can be determined.
The feature reduction operation can take various forms including one or more of: principal component analysis (PCA), random projection matrices (RPM), linear discriminant analysis, canonical correlation analysis, singular value decomposition (SVD), latent semantic indexing (LSI), discrete cosine transform (DCT), randomized principal component analysis (RPCA), or large sample randomized principal component analysis (LS-RPCA).
Non-transitory computer program products (i.e., physically embodied computer program products) are also described that store instructions, which when executed by one or more data processors of one or more computing systems, cause at least one data processor to perform operations herein. Similarly, computer systems are also described that may include one or more data processors and memory coupled to the one or more data processors. The memory may temporarily or permanently store instructions that cause at least one processor to perform one or more of the operations described herein. In addition, methods can be implemented by one or more data processors either within a single computing system or distributed among two or more computing systems. Such computing systems can be connected and can exchange data and/or commands or other instructions or the like via one or more connections, including but not limited to a connection over a network (e.g., the Internet, a wireless wide area network, a local area network, a wide area network, a wired network, or the like), via a direct connection between one or more of the multiple computing systems, etc.
The subject matter described herein provides many technical advantages. For example, the current subject matter provides enhanced techniques for preventing the bypass of machine learning models using techniques such as string stuffing which might otherwise cause such models to misclassifying analyzed artifacts. Further, the current subject matter is advantageous in that provides mitigation techniques can be used in connection with existing machine learning models (e.g., neural networks) without having to train and deploy new machine learning models which address model bypass techniques such as string stuffing.
The details of one or more variations of the subject matter described herein are set forth in the accompanying drawings and the description below. Other features and advantages of the subject matter described herein will be apparent from the description and drawings, and from the claims.
The current subject matter is directed to techniques for preventing techniques which seek to cause a classifier to misclassify an artifact. These techniques include, for example, string stuffing in which a plurality of strings (e.g., dozens, hundreds, thousands, etc.) are added to an artifact. The classifier in this regard can be an AI/machine learning model that outputs at least one value that characterizes the input to such model. While the current subject matter provides examples relating to models used for detection of malicious software (“malware”), it will be appreciated that the current subject matter can, unless otherwise specified, apply to other applications/workflows utilizing a model including, for example, autonomous vehicle navigation systems, image analysis systems, biometric security systems, video game cheat circumvention systems, and the like.
In some cases, the output of a classification model can be intercepted and exploited by a malicious actor as part of an adversarial attack. For example, data exchanged between a client and a remote server executing the classification model can be accessed such that small changes can be made to the data (e.g., file, code, artifact, etc.) input into the classification model until a desired outcome (from the point of view of the malicious actor) is obtained. For example, a malicious actor either automatically or through manual modifications can make small changes to a file encapsulating malicious code until such time that classification model determines that such file is safe to execute or otherwise access.
As will be described in further detail below, the resulting vectors as part of the feature reduction operation 140 can be used to generate an overflow vector 150. The overflow vector 150 can then be input into one or more classification models 160 (multiple model variations can sometimes be referred to as an ensemble of classification models 150) as well as the reduced feature vector (as part of operation 140).
The classification models 160 can take various forms including, without limitation, a logistic regression model, a neural network (including convolutional neural networks, recurrent neural networks, generative adversarial networks, etc.), a support vector machine, a random forest, a Bayesian model, and the like. The output of the classification models 160 can be a score 170. As used herein, unless otherwise specified, the score can be a numeric value, a classification type or cluster, or other alphanumeric output which, in turn, can be used by a consuming process 180 or application to take some subsequent action. For malware applications, the score can be used to determine whether or not to access, execute, continue to execute, quarantine, or take some other remedial action which would prevent a software and/or computing system from being infected or otherwise infiltrated by malicious code or other information encapsulated within the artifact 110.
In an arrangement in which the output of the model 160 can be readily ascertained, the score 160 can be used by a malicious actor to modify the artifact 195 and repeat the process until such time that a desired score 170 is output by the corresponding model 160. For example, the modified artifact 196 can encapsulate malicious script and small changes (i.e., addition of a plurality of strings) to the artifact 110 could result in the corresponding classification model 160 classifying such modified artifact 195 as being benign.
Modifications to an artifact 195 can be done in such a way as to maintain the original character or nature of the artifact 110. In the example of an actor attempting to modify a malicious file (malware), any modifications must be such that the malware still operates as intended. Such modifications can be made by (for instance) adding to, removing from, or altering un-used portions of the malicious file. As these portions of the file are unused, they have no effect on the realized behavior of the file, but may result in a different score 170 from the model 160. Alternatively or additionally, used sections of the artifact 110 can also be modified, so long as the final function of the malware is left intact.
Whether manually, or in an automated system, the actor or system will typically make many small changes, and get new scores 170 from the model 160. Any change that moved the score 170 in the desired direction (i.e. in the malware example, moving the score closer to a value that is interpreted as benign) is maintained, while other changes are discarded. Such an iterative process can be repeated until the cumulative changes to the artifact 110 result in a cumulative change in the score 170 which accomplishes the desired effect. The techniques provided herein can interrupt this cycle of iterative improvements by preventing such model manipulation from resulting in a false or misleading change in the score 170.
The vector modification operation 150 as used herein, modifies the contribution of a feature as it is added to the projected bucket to reduce the contribution of certain features which, when present or absent, provide a higher contribution to the score output by the classification model(s) 160. In one example, a projection matrix (at 140) can reduce the extracted features in the vector 130 into 10,000 positions (or buckets) as a means of dimensionality reduction prior to those features being presented to the model(s) 160. The current techniques address the situation in which a small number of such buckets can have a disproportionate contribution to the resulting score as shown in diagram 200 of
In one example, a portable executable (PE) model for malware analysis includes a feature set having a number of submodel outputs for autoit, nsis, entrypointopcode, and virtual ordering, as well as some heavily weighted features such as the PE checksum correct feature. In the Random Projection Matrix these submodel features are subject to random projection just as all other features. However, if you consider the graph of
One technique for modifying the artifact (at 195) in an effort to cause the model 160 to misclassify is to add (i.e., stuff) the artifact with numerous strings of data (e.g., nonsensical alphanumeric text strings, passages from websites, etc.). As will be noted in
As will be noted again with reference to
Rather than retrain the models 160 or to change the projected matrix mapping (as part of the feature reduction operation 140), with the current subject matter, the projected matrix utilized by the feature reduction operation 140) can be modified to: (i) reduce (i.e. attenuate) the contribution for the specific string features that land in the top 1% of buckets that most influence score towards positive, as defined by an aggregation of fuzzed malware baseline samples; and (ii) cause all non-string features landing in those buckets to retain their original (unmodified) contribution as they sum into the projected vector.
Such an arrangement is possible because the vectorization process can provide the ability to set the contribution of a feature projecting into a bucket by any value between +1.0 and −1.0.
The vector that represents the score influence for each bucket (aggregated over hundreds of malware samples) can be used to reduce the contribution of those string features that are assigned to the top 1% by a factor that levels their score contribution to be the same as other binary features at the 99-th percentile. This is specifically designed to flatten the left-most tail in
In one example, a disk controller 448 can interface with one or more optional disk drives to the system bus 404. These disk drives can be external or internal floppy disk drives such as 460, external or internal CD-ROM, CD-R, CD-RW or DVD, or solid state drives such as 452, or external or internal hard drives 456. As indicated previously, these various disk drives 452, 456, 460 and disk controllers are optional devices. The system bus 404 can also include at least one communication port 420 to allow for communication with external devices either physically connected to the computing system or available externally through a wired or wireless network. In some cases, the at least one communication port 430 includes or otherwise comprises a network interface.
To provide for interaction with a user, the subject matter described herein can be implemented on a computing device having a display device 430 (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information obtained from the bus 404 via a display interface 414 to the user and an input device 432 such as keyboard and/or a pointing device (e.g., a mouse or a trackball) and/or a touchscreen by which the user can provide input to the computer. Other kinds of input devices 432 can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback by way of a microphone 436, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input. The input device 432 and the microphone 436 can be coupled to and convey information via the bus 404 by way of an input device interface 428. Other computing devices, such as dedicated servers, can omit one or more of the display 440 and display interface 414, the input device 432, the microphone 436, and input device interface 428.
One or more aspects or features of the subject matter described herein can be realized in digital electronic circuitry, integrated circuitry, specially designed application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) computer hardware, firmware, software, and/or combinations thereof. These various aspects or features can include implementation in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device. The programmable system or computing system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.
These computer programs, which can also be referred to as programs, software, software applications, applications, components, or code, include machine instructions for a programmable processor, and can be implemented in a high-level procedural language, an object-oriented programming language, a functional programming language, a logical programming language, and/or in assembly/machine language. As used herein, the term “machine-readable medium” refers to any computer program product, apparatus and/or device, such as for example magnetic discs, optical disks, memory, and Programmable Logic Devices (PLDs), used to provide machine instructions and/or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term “machine-readable signal” refers to any signal used to provide machine instructions and/or data to a programmable processor. The machine-readable medium can store such machine instructions non-transitorily, such as for example as would a non-transient solid-state memory or a magnetic hard drive or any equivalent storage medium. The machine-readable medium can alternatively or additionally store such machine instructions in a transient manner, such as for example as would a processor cache or other random access memory associated with one or more physical processor cores.
To provide for interaction with a user, the subject matter described herein may be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) and/or a touch screen by which the user may provide input to the computer. Other kinds of devices may be used to provide for interaction with a user as well; for example, feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, speech, or tactile input.
In the descriptions above and in the claims, phrases such as “at least one of” or “one or more of” may occur followed by a conjunctive list of elements or features. The term “and/or” may also occur in a list of two or more elements or features. Unless otherwise implicitly or explicitly contradicted by the context in which it is used, such a phrase is intended to mean any of the listed elements or features individually or any of the recited elements or features in combination with any of the other recited elements or features. For example, the phrases “at least one of A and B;” “one or more of A and B;” and “A and/or B” are each intended to mean “A alone, B alone, or A and B together.” A similar interpretation is also intended for lists including three or more items. For example, the phrases “at least one of A, B, and C;” “one or more of A, B, and C;” and “A, B, and/or C” are each intended to mean “A alone, B alone, C alone, A and B together, A and C together, B and C together, or A and B and C together.” In addition, use of the term “based on,” above and in the claims is intended to mean, “based at least in part on,” such that an unrecited feature or element is also permissible.
The subject matter described herein can be embodied in systems, apparatus, methods, and/or articles depending on the desired configuration. The implementations set forth in the foregoing description do not represent all implementations consistent with the subject matter described herein. Instead, they are merely some examples consistent with aspects related to the described subject matter. Although a few variations have been described in detail above, other modifications or additions are possible. In particular, further features and/or variations can be provided in addition to those set forth herein. For example, the implementations described above can be directed to various combinations and subcombinations of the disclosed features and/or combinations and subcombinations of several further features disclosed above. In addition, the logic flows depicted in the accompanying figures and/or described herein do not necessarily require the particular order shown, or sequential order, to achieve desirable results. Other implementations may be within the scope of the following claims.
