Patent Application
20230031052
The present invention relates generally to federated learning in computer systems. Methods for model-based federated learning are provided, together with computer systems implementing such methods.
Federated Learning (FL) refers generally to machine learning techniques in which a set of participants cooperate in a machine learning process in order to benefit from heterogeneous, often geographically dispersed, data available to the individual participants. Machine learning (ML) is a cognitive computing technique in which a dataset of training samples from some real-world application is processed in relation to a basic model for the application in order to train, or optimize, the model for the application in question. After learning from the training data, the trained model can be applied to perform inference tasks based on new (previously unseen) data samples for the application. ML techniques are used in numerous applications in science and technology, including medical diagnosis, image analysis, speech recognition/natural language processing, genetic analysis and pharmaceutical drug design, among a great many others.
Performance of ML models is highly dependent on the size and diversity of the training datasets. However, movement of data is increasingly restricted by data privacy regulations and security issues, inhibiting distribution of data for training ML models. This is a significant problem where distributed parties, each with their own silo of training data, wish to cooperate and benefit from each other's training data. FL provides techniques to address such issues.
Conventional FL provides a distributed learning process in which the participating computers (i.e., node computers), each with a local training data silo, can interact to build a common, robust ML model without sharing their local training data. During training, updates to the parameters of local models, trained on local datasets, are aggregated to produce a global model which is then distributed to all nodes for further training. For example, IBM® Federated Learning (IBM FL) (IBM and all IBM-based trademarks and logos are trademarks or registered trademarks of International Business Machines Corp. and/or its affiliates) provides state-of-the-art protocols for enterprise-grade federated learning, with plug-ins for enhancing privacy and security, such as differential privacy and secure multi-party communication. In some scenarios, however, it may not be possible or desirable for parties to build a common, shared model and/or ML models may need to be deployed at resource-constrained devices where the computationally intensive training of models is infeasible.
According to an embodiment, the present invention provides a method for federated learning among a federation of machine learning models in a computer system. The method includes, in at least one node computer of the system, deploying a federation model for inference on local input data samples at that node computer to obtain an inference output for each data sample, and providing the inference outputs for use as inference results at that node computer. The method further comprises, in the system, for at least some of the data samples, obtaining an inference output corresponding to each data sample from each of at least a subset of the other federation models, and using the inference outputs from the federation models to provide a standardized inference output corresponding to an input data sample at the node computer for assessing performance of the model deployed at that computer.
Also, according to an embodiment, the invention provides a computer system for implementing a federated learning method described above.
Embodiments of the invention offer model-based FL methods/systems in which performance of a pre-trained federation model, which is actively deployed for inference on local data samples at a node computer, can be assessed using a standardized inference output for those samples. The standardized inference output, which can be produced in various ways explained below, is obtained by using inference outputs from other models in the federation, and thus provides a federation-based standard for assessing inference results at a given node. Inference results for data samples at a node computer can be assessed on a sample-by-sample basis. This provides a basis for various actions, detailed below, to be taken to ensure appropriate performance at a node computer and to share learning between the federation models.
Embodiments can be implemented with pre-trained models, permitting use with node computers in which training of models is restricted or infeasible, without requiring access to the original training data. For example, node computers of the system may comprise edge devices in a data communications network. Such edge devices may, for example, comprise mobile phones, personal computing devices, IoT (Internet of Things) sensors or other IoT devices which may have limited compute resources and/or need to function offline where necessary.
Different federation models may be deployed for inference at different node computers, while maintaining a required performance standard throughout. Embodiments can address scenarios in which different parties wish to maintain security of the parties' own ML models while still benefiting from each other's learning. For example, competing companies may wish to mutually benefit from each other's learning based on different training datasets, without sharing those datasets or their local models. Embodiments can also address scenarios in which multiple parties need to ensure comparable model predictions while preserving data confidentiality. For example, a consortium of banks may seek to establish a multi-model performance benchmark for particular applications such as loan approval or credit risk scoring. In such cases, each party may locally train and deploy a federation model for inference at a node computer of the system, while inference results at each node can be assessed on a sample-by-sample basis in relation to a federation standard.
In some embodiments, node computers may communicate directly with other federation nodes via a data communications network. In other embodiments, the system may include a control server for communication with the node computers via a data communications network. The method may then include, at each node computer, sending to the control server inference data defining an input data sample and the inference output for that data sample at that node computer, and, at the control server, using the inference data to request an inference output corresponding to that data sample from each of at least a subset of the federation models at other node computers. The control server can then use the inference outputs from the federation models to provide a standardized inference output corresponding to an input data sample at each node computer. The control server may be implemented here by a trusted entity/regulatory authority in some embodiments. In either communications scenario, communications can be implemented in a confidential computing environment where required, such that security of confidential information is protected in operation of the system.
The standardized inference output corresponding to an input data sample may be produced as a function of the inference outputs from the federation models for that sample. As examples here, the standardized inference output may comprise one of a majority vote and an average derived from the inference outputs for a data sample. This provides a particularly simple implementation which also inhibits so-called “poisoning” of the system as discussed further below. Standardized outputs may also exploit confidence values associated with the inference outputs, where available, as illustrated by embodiments below.
Further advantageous embodiments include, at least in a preliminary operating phase of the system, using the inference outputs from the federation models corresponding to each data sample to train a further ML model, or “metamodel”, which is then included in the federation. After training the metamodel, an inference output for an input data sample may be obtained from (at least) the metamodel to provide the standardized output corresponding to that sample. By using the inference outputs of federation models for metamodel training, performance of the metamodel can be expected to exceed that of any individual model in the federation, providing a convenient federation-wide standard for assessment of all models. For example, the aforementioned control server may alert a node computer if its local inference output deviates in a predetermined manner from the standardized output. This provides an elegant system for benchmarking/regulation of federation models, e.g., in banking/insurance/financial or healthcare scenarios where mutually-consistent inference results can be critical for a federation.
Alternative embodiments may include, at a node computer deploying a federation model for inference, storing at least a subset of the other federation models, and obtaining inference outputs from each of the other stored models for local input data samples at the node computer. The inference outputs from those other models can then be used to produce the standardized inference output corresponding to each input data sample at the node computer. Here, federation nodes can use one model for active inference at that node, with other federation models operating in a “shadow mode” for obtaining a standardized output for each local inference sample. Embodiments here may also exploit features of other embodiments above, such as training and use of metamodels. Metamodels can be advantageously deployed as “challengers” to federation models in some embodiments. These and other features and advantages will be described in relation to particular embodiments below.
Embodiments of the invention will be described in more detail below, by way of illustrative and non-limiting example, with reference to the accompanying drawings.
Each ML model 3 is pretrained, either locally or prior to provision in a node computer 2, and is deployed for inference on local input data samples at the node computer. The nature of the input data samples, and the particular inference task performed, depends on the nature and function of the federation in question. ML-based inference generally falls into one of two categories, namely classification or regression. Classification tasks assign input data samples to one of a discrete set of predefined categories, or classes, and the model output for a given input sample indicates the particular class to which that sample is assigned. Regression tasks generally output a value (or value range) for some predefined continuous variable based on processing of an input sample by the model. Numerous types of federations and inference applications can be envisaged for implementation in system 1. As illustrative examples only, models may be deployed for tasks such as: image classification, e.g. for identifying particular subject matter in digital images or digital video; audio analysis, e.g. for speech recognition tasks; medical diagnosis, e.g. for classifying pathology images as diseased/healthy or evaluating severity of cancer tumors by regression analysis of tumor slides; text processing tasks, e.g. predictive text for user input devices; banking/business applications, e.g. evaluating risk for loan applications, approving insurance policies, or identifying/qualifying faults in structures in the building industry; and pharmaceutical drug selection, e.g. predicting efficacy of drugs for treatment of specific patients. Numerous other applications in technical, commercial, industrial and healthcare settings can also be envisaged.
ML models 3 may comprise any type of ML model as appropriate for the required inference task. Numerous ML models are known in the art, such as neural networks (including deep neural networks), tree-ensemble models (such as Random Forests models), Bayesian networks, SVMs (Support Vector Machines), and so on. Suitable models may be selected as appropriate for a required inference task. Note also that different models (or types of models) can be employed at different node computers where different models can perform the inference task in question.
In some applications, a node computer may comprise a general-purpose user computer such as a desktop, laptop or tablet computer. Node computers may also comprise mobile phones, smart speakers, televisions, personal music players or other such user devices. Node computers may further comprise sensors or other devices in the Internet of Things. In general, however, node computers 2 may be implemented by any type of general- or special-purpose computer, which may comprise one or more (real or virtual) machines, providing functionality for implementing the operations described herein. Federation control server 5, where provided, may similarly be operated by one or more (real or virtual) machines providing server functionality for managing operation of node computers in the federation. Such a control server may be implemented by a party running or controlling a given federation, e.g., as web server or a server operated by a regulatory authority or trusted entity for the federation. Computers 2, 5 in system 1 may also be implemented in a distributed cloud computing environments where tasks are performed by distributed processing devices linked via a communications network.
The block diagram of
Bus 13 represents one or more of any of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, and a processor or local bus using any of a variety of bus architectures. By way of example, and not limitation, such architectures include Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.
Computer 10 typically includes a variety of computer readable media. Such media may be any available media that is accessible by computer 10 including volatile and non-volatile media, and removable and non-removable media. For example, system memory 12 can include computer readable media in the form of volatile memory, such as random access memory (RAM) 14 and/or cache memory 15. Computer 10 may further include other removable/non-removable, volatile/non-volatile computer system storage media. By way of example only, storage system 16 can be provided for reading from and writing to a non-removable, non-volatile magnetic medium (commonly called a “hard drive”). Although not shown, a magnetic disk drive for reading from and writing to a removable, non-volatile magnetic disk (e.g., a “floppy disk”), and an optical disk drive for reading from or writing to a removable, non-volatile optical disk such as a CD-ROM, DVD-ROM or other optical media can also be provided. In such instances, each can be connected to bus 13 by one or more data media interfaces.
Memory 12 may include at least one program product having one or more program modules to carry out functions of embodiments of the invention. By way of example, program/utility 17, having a set (at least one) of program modules 18, may be stored in memory 12, as well as an operating system, one or more application programs, other program modules, and program data. Each of the operating system, one or more application programs, other program modules, and program data, or some combination thereof, may include an implementation of a networking environment. Program modules 18 may generally carry out functions and/or methodologies of embodiments of the invention as described herein.
Computer 10 may also communicate with: one or more external devices 19 such as a keyboard, a pointing device, a display 20, etc.; one or more devices that enable a user to interact with computer 10; and/or any devices (e.g., network card, modem, etc.) that enable computer 10 to communicate with one or more other computing devices. Such communication can occur via Input/Output (I/O) interfaces 21. Also, computer 10 can communicate with one or more networks such as a local area network (LAN), a general wide area network (WAN), and/or a public network (e.g., the Internet) via network adapter 22. As depicted, network adapter 22 communicates with the other components of computer 10 via bus 13. Computer 10 may also communicate with additional processing apparatus 23, such as one or more GPUs (graphics processing units), FPGAs, or integrated circuits (ICs), for implementing functionality of embodiments of the invention. It should be understood that although not shown, other hardware and/or software components could be used in conjunction with computer 10. Examples include, but are not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems and data archival storage systems, etc.
Basic steps of model-based FL methods embodying the invention are indicated in
Steps 31 to 35 of the
When model 3 is deployed for inference, inference module 44 receives data samples for which inference is to be performed from one or more local applications 47 at the node computer. Each input sample is supplied to the model (typically in the form of a “feature vector” which represents the sample in a predetermined format used for model inputs during training and is generated by inference model 44 for the sample), to obtain the inference output, e.g., a classification, for the sample. The inference output is then returned to local application 47 as the inference result for the data sample and may be output to a user or otherwise used by application 47 depending on the use scenario. In addition, for at least some data samples processed by model 3, model controller 42 provides the sample (or feature vector) and the inference output for that sample to FL controller 43. The FL controller provides functionality for communication with control server 5 in this embodiment. In particular, FL controller implements the necessary communications protocols for communicating with server 5, and can also implement security protocols (e.g., data privacy and/or encryption protocols) for ensuring confidentiality of communications to the extent required in the federation system.
Functionality of logic modules 42 through 45 may be implemented, in general, by software (e.g., program modules) or hardware or a combination thereof. Functionality described may be allocated differently between system modules in other embodiments, and functionality of one or more modules may be combined.
In step 53, control server 5 uses the inference outputs from the federation models to produce a standardized inference output Sout corresponding to the input data sample in question. This standardized output Sout can be produced as a function of all the inference outputs from the federation models for the sample. Various functions can be envisaged here. For classification models, for example, Sout may be determined by a majority vote among the classification outputs of the various models. For regression models, Sout may be calculated as an average (e.g., a mean) derived from the values output by the models. Where federation models indicate a confidence value associated with an inference output (as is typically the case for ML models), determination of Sout may depend on the confidence values associated with the model outputs. For example, only outputs above a threshold confidence level may be used and/or regression values may be weighted by confidence to obtain Sout as a weighted average. A confidence value for Sout itself may also be calculated, e.g., as an average of the confidence values for the contributing model outputs.
In step 54, control server 5 assesses performance of the model at the node which sent the inference data (step 51) in relation to the standardized output Sout. In this embodiment, the control server checks whether the model output, as defined by the received inference data, deviates in a predetermined manner from Sout, and alerts the node computer if so. Various alert criteria may be defined here, e.g., that the model output corresponds to a different classification to Sout, a regression output deviates by more than a threshold amount from Sout, or the confidence value for the model output differs by more than a threshold amount from that calculated for Sout. Suitable alert criteria can be defined as desired for a given federation task.
An alert may be handled in various ways at a node computer 2. Control server 5 may send Sout to the node computer, and module 45 may adjust parameters of local model 3 accordingly. For example, module 45 may use Sout as a training label for the input sample in a training stage for the model or may otherwise adjust the local model parameters so as to mitigate deviation of the model output from Sout.
The
In some embodiments, steps 51 to 54 of
The above systems provide effective techniques for multi-model monitoring and benchmarking in federations of models, enabling comparable model performance to be ensured across a federation. Benchmarking is important in numerous application scenarios to ensure mutually-consistent performance of different federation models. In the healthcare industry, for example, it can be critical for private models at different institutions to produce consistent results. Regulation in other industries, such as banking and other financial, commercial or industrial applications, often requires distributed models to meet industry performance benchmarks. Moreover, models at individual nodes can be improved based on better-performing models at other nodes, allowing models to benefit from each other's learning. In addition, by assessing model performance using a standardized output derived from a plurality of federation models, the system is protected from so-called poisoning by any one federation model. If one federation node (intentionally or otherwise) injects bad results into the system, this will be mitigated by the standardization process.
While a federation control server 5 is provided in embodiments above, systems can be envisaged in which nodes can communicate directly with other federation nodes. Operations performed by the control server above may be implemented by individual federation nodes in these systems. For example, nodes may include local functionality for generating a standardized output Sout for their inference samples.
Another implementation of a model-based FL method will now be described with reference to
Node computer 58 may compare the standardized output Sout with the inference output of the active model, and if the active model output deviates in a predetermined manner from Sout, the node computer may adjust parameters of the active model to alleviate the deviation. For example, the active model may be further trained based on the inference outputs from the shadow models, e.g., using Sout as a training label for the data samples.
It will be seen that the above embodiments offer highly-effective systems for model-based federated learning. However, various alternatives and modifications can be made to the particular embodiments described. By way of example, features described with reference to one embodiment may be applied in other embodiments as appropriate. In general, where features are described herein with reference to a method embodying the invention, corresponding features may be provided in a system embodying the invention, and vice versa.
The descriptions of the various embodiments of the present invention have been presented for purposes of illustration but are not intended to be exhaustive or limited to the embodiments 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 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.
