A method for machine unlearning in federated learning

Abstract

There is provided a computer-implemented method for machine unlearning in federated learning The method comprises obtaining a global model. The global model is constructed using contribution weights and disentangled model parameters associated with a plurality of agents. The method further comprises obtaining an unlearn request, the unlearn request indicating a first agent of the plurality of agents, identifying a first contribution weight for a first set of disentangled model parameters associated with the first agent, and reducing the identified first contribution weight. The method further comprises updating the global model based on the reduced first contribution weight, generating a message comprising the updated global model, and transmitting the message towards a second agent.

Description

A METHOD FOR MACHINE UNLEARNING IN FEDERATED LEARNING TECHNICAL FIELD

[0001] Disclosed are embodiments related to federated learning, federated unlearning, machine unlearning, and determining contributions or influence of agents. BACKGROUND

[0002] Machine learning operations (“MLOps”) consists of many sub-operations performed for machine learning (“ML”) model maintenance starting from its initialization, training, hyper- parameter tuning, and then further adapting to changes (e.g., data drift, concept drift) to sustain good model efficacy.

[0003] After a model is trained, a re-training step may be warranted when a subset of data that was used in training needs to be removed. A subset of data may need to be removed from the training dataset for several reasons, such as due to model performance drop (data drift, concept drift), change in environmental, regional or governmental regulations, the data containing sensitive or private information (e.g., may be including sensitive data from end-users), or changes in business related incentives for training a joint model with potentially competitive entities. In the context of federated learning, one may want to remove contributions or influence of one or more agents due to the aforementioned reasons or the presence of one or more agents that are detected to be intentionally or unintentionally impacting the overall model performance negatively.

[0004] In scenarios where a portion of a dataset that was used for training needs to be removed, the ML model needs to be retrained in order to obtain a new version of the previous model without revealing potential information learned from the portion of the dataset that was asked to be removed. This scenario is known as machine unlearning. In some cases, training takes a long time (depending on the available hardware resources, and the size of model and dataset), which in parallel increases the storage, memory, computation, and energy supply requirements. Even if the computation resources are available, in some cases training datasets may no longer be accessible. For example, in some instances data must be deleted after a certain time period due to, for example, privacy regulations. Accordingly, it is desirable to avoid retraining a ML model from scratch after the removal of a portion of the dataset in order to, for example, reduce operational costs such as memory, computation, network footprint, and energy consumption in MLOps.

[0005] In telecommunications, machine unlearning may be a required feature due to increasing levels of liability and corresponding strictness in data protection regulations. For example, a model should contain contributions only from clients that are willing to share and collaborate.. In the case of collaborative training such as federated learning, the ownership of a ML model is not well- defined, and it is non-trivial to pinpoint underlying reasons if a ML model under-performs than expected. Similarly, a base station node may consist of different network elements from different vendors, e.g., antenna and radio units from one vendor, while the baseband and power supplier unit from another. A joint model trained on dataset collected from products of different vendors may also create similar challenges with respect to business sensitivity as well as model ownership and responsibility.

[0006] There are distributed learning techniques that use a centralized or dedicated server to maintain global parameters, such as Federated Learning (“FL”), that address the privacy sensitive aspects up to some extent. However FL execution is solely based on FedAvg operation (iterative averaging of clients’ model parameters and continual training). Such mechanisms are trained over plurality of training iterations, i.e., averaging of multiple rounds of contributions received from multiple clients. The iterative training of aggregated, and therefore, entangled contributions throughout the life time of a federated training makes disentanglement of contributions received from multiple clients even harder. SUMMARY

[0007] Unlearning under FL setups can be inherently challenging due to the entangled contributions of each client during successive federation rounds. This complexity arises from the nature of FL, where the information from local models is first aggregated at the server’s end and then exchanged among participants. Entanglement refers to the integration of information from local models during the aggregation process. Disentanglement refers to separating out a contribution of individual data samples or local models from the global model.

[0008] In one category of methods [1, 2], an unlearning algorithm is implemented at the agent side, motivated by the fact that in the context of federated learning, the server entity lacks access to training data. However, these methods bring potential reliability and security risks to a FL system. This is largely due to the commonly recognized assumption in FL that agents may manipulate their data or model weights to their advantage, thereby creating added challenges tomaintaining secure and efficient unlearning in a federated environment. Additionally, these methods need to have access to the original data used for the training, which is a limiting assumption because original data may not be available in many realistic scenarios.

[0009] FedEraser [3] constructs the unlearning algorithm at the server. This technique focuses on minimizing the loss function of the global model while retaining the accuracy of unlearning below a specific threshold. Model parameters are calibrated using the Frank-Wolfe algorithm to solve an optimization problem, enabling the adjustment of the remaining agents' weights to offset the influence of the unlearned agent. One limitation of this technique is that the unlearning process itself requires optimization. The second limitation is that it requires having access to the original data.

[0010] Aspects of the present disclosure include a contribution model that determines a contribution (or sometimes referred to as “influence”) of the agents during a federated learning procedure. Contributions are computed in a global sense. In order to compute the contributions, the contribution model enforces a mechanism that encourages disentanglement of the contributions from each agent. The contribution model is located at the server node, and it is learned successively using the incoming model parameters from the agents at each round of federation. Accordingly, a new formulation of the FedAvg algorithm is proposed wherein the global model is a weighted average of the local models; the weights correspond to the global contribution of the agents. Once the contributions of the agents are determined, the task of unlearning one or more agents becomes reducing their contributions. In some embodiments, reducing an agent’s contribution may include setting their contribution to zero or to a predetermined or user-defined value.

[0011] According to one aspect, a computer-implemented method for machine unlearning in federated learning is provided. The method includes obtaining a global model, wherein the global model is constructed using contribution weights and disentangled model parameters associated with a plurality of agents. The method includes obtaining an unlearn request, the unlearn request indicating a first agent of the plurality of agents. The method includes identifying a first contribution weight for a first set of disentangled model parameters associated with the first agent. The method includes reducing the identified first contribution weight. The method includes updating the global model based on the reduced first contribution weight. The method includes generating a message comprising the updated global model. The method includes transmitting the message towards a second agent.

[0012] In another aspect there is provided a node with processing circuitry adapted to perform the method described above. In another aspect there is provided a computer program comprising instructions which when executed by processing circuity of a node causes the node to perform the methods described above. In another aspect there is provided a carrier containing the computer program, where the carrier is one of an electronic signal, an optical signal, a radio signal, and a computer readable storage medium. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate various embodiments.

[0014] FIG.1 is a flow diagram, according to some embodiments.

[0015] FIG.2 is a conceptual diagram, according to some embodiments.

[0016] FIG.3 illustrates error bar plots comparing performance of different learning frameworks, according to some embodiments.

[0017] FIG.4 is a flow diagram, according to some embodiments.

[0018] FIG.5 is a method, according to some embodiments.

[0019] FIG.6 a block diagram of a node, according to some embodiments. DETAILED DESCRIPTION

[0020] Aspects of the present disclosure include a method of unlearning in the context of federated learning where the contribution from each agent is learned in a global sense during the course of federation. A contribution model is constructed that takes in as inputs the set of local model parameters from all agents and produces a version where the model parameters are disentangled. During this process of disentanglement, the contribution model learns in a global sense how much each individual agent has contributed to the federated learning. During the course of federated learning, the global model is constructed by taking weighted average of the local models where the weights are the contributions from the agents. Then the federated unlearning is done by simply setting the contribution of the agent(s) to zero.

[0021] Unlike the current unlearning solutions [3], aspects of the techniques disclosed herein neither require optimization at the unlearning phase nor having access to the original data. As abyproduct of this solution, the contributions of each local model in the global model can be identified. However, in some embodiments, the contribution may not state whether the contribution from a local model is positive or negative. The techniques disclosed herein allow for removal of multiple agents / local models at one operation.

[0022] FIG. 1 is a flow diagram, according to some embodiments. Aspects of the present disclosure may include the following steps as shown in FIG.1. As part of initialization, a server sends the initial global model to the agents. In some embodiments, the global model is initialized randomly. In some embodiments, the global model is initialized with bootstrapped parameters, or parameters already learned. At the agent node, given the global model, each agent trains a local model using its local data.

[0023] At 101, each agent sends its local model to the server. The local model may include a set of model parameters for each agent. In some embodiments, each agent may send multiple copies of its local model to the server. Accordingly, each agent may run through multiple iterations of training and updating its local model, and send one or more of the versions of its local model to the server per round of updates to the global model at the server.

[0024] At 103, the contribution model receives the local model parameters as input. It then encodes the parameters into a matrix (e.g., a row or column matrix) of latent variables named latent contributions where each column represents latent contribution vector of an agent. Thus, the number of columns is equal to the number of agents. The length of each contribution vector (or “embedding”) may be a user defined parameter. The latent contributions are initially entangled. In some embodiments, the length of each contribution vector may be greater than the number of agents times 2 (2*n_agents). Accordingly, if there are five agents, the length of each contribution vector would be at least ten. In some embodiments, depending on the size of the model, the length of the embeddings may be increased. In some embodiments, an optimal value for the embedding size may be selected through a hyper-parameter search.

[0025] At 105, the latent contributions are output by the encoder function. The latent contributions may be output as a matrix of (tangled) latent contribution vectors where each column represents an agent latent contribution vector. The latent contribution vectors may be tangled by nature.

[0026] At 107, each latent contribution vector then passed through a disentanglement functionality that produces a modified version of the latent contribution vector that is disentangled from other latent contribution vectors. As an example, this disentanglement can be achieved by enforcingorthogonality across latent contributions by adding regularization terms in the objective loss function, which encourage the latent contributions to become orthogonal.

[0027] At 109, a set of disentangled latent contributions is output. In some embodiments, the disentangled latent contributions is a matrix of disentangled latent contribution vectors.

[0028] At 111, the sum of the elements of a disentangled latent contribution vector is a scalar quantity that indicates the contribution of an agent, named contribution weights.

[0029] At 113, the disentangled latent contributions are passed through a decoder that takes them as the input.

[0030] At 115, the decoder produces as the output a reconstructed version of the model parameters that are disentangled, named disentangled model parameters.

[0031] At 117, given the disentangled model parameters and the contribution weights, a global model is constructed by taking an average of the disentangled model parameters weighted by the contribution weights.

[0032] At 119, the global model is output. In some embodiments, the global model comprises aggregated global model parameters.

[0033] At 121, the server sends the global model to the agents.

[0034] In some embodiments, the steps shown in FIG. 1 are repeated until a convergence criterion is met.

[0035] Notation for aspects of the present disclosure is described below.

[0036] Let ^^ denote the set of agents. For the ease of discussion, consider a federation of three agents, ^^ ={^^, ^^, ^^}. It should be appreciated that a federated learning environment may include significantly more than three agents. Let ^^^^= {( ^^^^^^, ^^^^^^)| ^^ = 1,2, … } denote the training data for agent "a" containing the pairs of inputs ^^^^^^and outputs ^^^^^^; similarly assume ^^^^= {( ^^^^^^, ^^^^^^)| ^^ = 1,2, … } and ^^^^= {( ^^^^^^, ^^^^^^)| ^^ = 1,2, … } for agents "b" and "c", respectively.

[0037] Predictive model.

[0038] Let ^^( ^^^^; ^^^^) denote the predictive model for agent "a", where ^^^^denotes all the model parameters; similarly assume ^^( ^^^^; ^^^^) and ^^( ^^^^; ^^^^) for agents "b" and "c", respectively.

[0039] Contribution model.

[0040] Let ^^ ≔ vec( ^^^^, ^^^^, ^^^^) be a vector representation of all the parameters from all agents and let ^^ ≔ vec( ^^^^, ^^^^, ^^^^) denote the concatenation of the contribution latent variablesfrom all agents. In some embodiments, the latent contributions ^^ and the model parameters ^^ are not directly comparable. Thus, a second group of latent variables may be defined, named latent support set that projects a latent contribution into a space where it can be compared against the model parameters. As an example, for the agent "a" with the latent contribution ^^^^and model parameters ^^^^, the latent support set is denoted as ^^^^≔ { ^^^^^^, ^^^^^^}. The product of the support variables ^^^^and the latent contributions ^^ are referred to as the projected model parameters defined as ^^̃^^^^= ^^^^( ^^^T^^^^^^^) and as ^^̃^^^^= ^^^^( ^^^T^^^^^^^) which can now be compared against the agent model parameters ^^^^. Similarly for agent "b", the latent support set is denoted as ^^^^≔ { ^^^^^^, ^^^^^^} for which the projected model parameters are defined as, ^^̃^^^^= ^^^^( ^^^T^ ^^^^^^) and ^^̃^^^^= ^^^^( ^^^T^ ^^^^^^). Finally, for agent "c", the latent support set is denoted as ^^^^≔ { ^^^^^^, ^^^^^^} for which the model parameters are defined as: ^^̃^^^^= ^^^^( ^^^T^^^^^^^) and ^^̃^^^^ = ^^^^( ^^^T^^^^^^^).

[0041] The model is constructed as follows as shown in equation (1):learnable parameters ^^encare the inner encoder and decoder models with the learnable aagent "a". The inner encoders and decoders are responsible for generation of the latent contribution variables ^^ through the latent variables ^^^^, ^^^^, and ^^^^.

[0042] FIG. 2 is a conceptual diagram, according to some embodiments. FIG. 2 visualizes the relation between the variables described herein, showing dependencies between model parameters and the latent contribution variables. The model parameters ^^^^, ^^^^, and ^^^^for each of the three agents are shown at 201. The latent contributions for each agent ^^^^, ^^^^, and ^^^^are shown at 203. Each agent model parameter is represented by a vector that needs to be learned. The latent contributions are to be learned such that they are disentangled from each other, as shownin 205, 207, 209, and 211. The model receives the agents’ model parameters and learns the contributions from the agents.

[0043] The model parameters at 205 are processed by an outer encoder model ^^^^enc. The latent contributions ^^ (209) are output along with the latent support sets ^^^^, ^^^^, ^^^^

[0044] The inner decoder models ^^^^e^^nc, ^^^^^e^nc, and ^^^^^e^ncare shown at 213, which generate the latent contribution variables ^^ , shown at 219, through the latent variables ^^^^, ^^^^, and ^^^^, shown at 215. At 217, the inner decoder models ^^^^d^^ec, ^^^^d^^ ec, and ^^^^d^^ecare shown. The latent contribution variables are processed by ^^^^dec, the outer encoder, shown at 221 to produce the global model shown at 223.

[0045] Training of the contribution model at the server.

[0046] In some embodiments, the training is carried out via minimization of the objective loss function through a procedure of training neural networks via error backpropagation. The optimal settings of the learnable parameters may be given by the settings of the parameters that minimize the loss, that is: , where ^^ is the optimization loss

[0047] Optimization loss: In some embodiments, latent contributions from the agents are learned such that they are generated from disentangled processes. Taking agent "a" as an example, it means ^^^^⊥ ^^^^and ^^^^⊥ ^^^^, where ⊥ denotes the disentanglement symbol. Similarly for agent "b", one may enforce ^^^^⊥ ^^^^and ^^^^⊥ ^^^^, and for agent "c", one may enforce ^^^^⊥ ^^^^and ^^^^⊥ ^^^^. Furthermore, one may further require disentanglement of a given agent model parameters from its projected model parameters. Taking agent "a" as an example, that means ^^^^⊥ ^^̃^^^^and ^^^^⊥ ^^̃^^^^. Similarly for agent "b", one may enforce ^^^^⊥ ^^̃^^^^and ^^^^⊥ ^^̃^^^^, and for agent "c", one may wish to enforce ^^^^⊥ ^^̃^^^^and ^^^^⊥ ^^̃^^^^.

[0048] Aspects of the present disclosure propose the following loss function that encourages disentanglement across the latent contribution variables and in between the model parameters and their projected model parameters. The loss function may defined as shown in equation (2):

[0049] where ℓrecand ℓregare reconstruction loss and regularization loss functions. The reconstruction loss functions may be defined as shown in equations (3) and (4):

[0050] The regularization losses may be designed to encourage disentanglement among the variables. The first regularization loss, for agent "a" as an example, is designed to encourage ^^^^⊥ ^^̃^^^^and ^^^^⊥ ^^̃^^^^, defined as shown in equations (5a) and (5b):function. In some embodiments, similarity functions other than the cosine function may alternatively be used. Similarly, the second regularization loss, for agent "a" as an example, is designed to encourage ^^^^⊥ ^^^^and ^^^^⊥ ^^^^, as shown in equation (6):model at the server (forward computation).

[0053] In some embodiments, the inference involves a single forward pass through the model described in Equation (1). Given a set of entangled model parameters ^^ ≔ vec( ^^^^, ^^^^, ^^^^), the contribution model produces a disentangled version of these parameters ^^̂ ≔ vec( ^^̂^^, ^^̂^^, ^^̂^^), and their corresponding disentangled contribution latent variables ^^̂ ≔ vec( ^^̂^^, ^^̂^^, ^^̂^^). Given the disentangled contribution latent variables, one may compute the contribution weights from each agent, ^^^^= ∑ ^^̂^^, ^^^^= ∑ ^^̂^^, and ^^^^= ∑ ^^̂^^.

[0054] Weighted Aggregation.

[0055] Given the inferred weight contributions from the agents and the set of disentangled model parameters, the aggregated model is computed as shown in equation (7):

[0057] Given that the contribution model has identified global contribution weights from participating agents, the unlearning task would become setting to zero the contribution weights of an agent that one would like to unlearn. Taking as an example agent "c" as the agent one would like to unlearn, the aggregated model in Equation (7) becomes:Table 1 shows a learning to unlearn algorithm, according to some embodiments. Table 1

[0059] In some embodiments, the algorithm may be divided into a learning part and an unlearning part. An example of such algorithm is provided below.

[0060] Example application. Set forth below is an example application, according to some embodiments.

[0061] Dataset: A MNIST dataset for the digit (0-9) classification task is used. Table 2 illustrates an assignment of training data to 5 agents under uneven distribution.Table 2

[0062] Methods: the performance of the following methods is compared: (1) FLAvg: Standard federated learning using average as the aggregation function. (2) FLAvg (subset): Federated learning with agents retrained from the beginning. (3) FLZ: the proposed federated learning-unlearning framework described herein. (4) FLZ (unlearn): The unlearned version of the FLZ model disclosed herein, aligned with the proposed unlearning mechanism. (5) LL: Local learning.

[0063] Given an agent to unlearn, FLAvg (subset) corresponds to a case where one has removed the agent and retrained the model. Thus, it should be seen as an upper bound on the performance. In comparison, FLZ (unlearn) carries out the unlearning task without the need for retraining of the model. For the case of FLAvg (subset) and FLZ (unlearn), agent 4 is removed from the federation and its effect is unlearned from the model.

[0064] FIG. 3 illustrates error bar plots comparing performance of the different learning frameworks, according to some embodiments. Classification results were evaluated using a metric of accuracy and are shown in FIG.3. As shown in FIG.3, removing the agent from the federation has helped FLAvg(subset) to achieve a performance better than FLAvg. This means that the agent had a negative effect on the federation. A similar trend is observed for FLZ and FLZ (unlearn). Comparing directly FLZ (unlearn) with FLAvg (subset), FLZ (unlearn) has a lower performance than the upper bound, FLAvg (subset). However, the gap is marginal given the fact that FLZ (unlearn) does not require training.

[0065] In Federated unlearning setting, the computation at the server end may be expensive due to increased number of model parameters to learn simultaneously, as it is responsible for learning the embeddings, contributions during training. Due this reason, some of the computation tasks may be offloaded to other available computation units in the cloud.

[0066] FIG. 4 is a flow diagram, according to some embodiments. FIG. 4 illustrates a potential implementation of aspects of the present disclosure in O-RAN, which may require slight extension of computation units and communication protocols. FIG. 4 illustrates a message flowbetween a collector (401), non-RT RIC (403), Near-RT RIC (405), O-CU / O-DU-1 (407), O- CU / O-DU-2 (409), O-CU / O-DU-N (411), Application Server 1 (413), Application Server 2 (415), and Application Server N (417). At steps 419, data for collection is sent from Application Server 1 (413), Application Server 2 (415), and Application Server N (417) to O-CU / O-DU-1 (407), O- CU / O-DU-2 (409), O-CU / O-DU-N (411), respectively. At steps 421, O-CU / O-DU-1 (407), O- CU / O-DU-2 (409), O-CU / O-DU-N (411) train their local models based on the data collection. At steps 423, the local model parameters are transmitted to the collector (401), and the non-RT RIC (403) obtains the collected models from the collector (401).

[0067] At steps 425, aspects of the present techniques are performed. The ML models are concatenated, the embeddings are learned, the contributions are learned, the embeddings are disentangled, the disentangled weights are determined, the loss function is computed, the loss is minimized, and the global model is updated. Additionally, an unlearn request may be transmitted from an application server to a respective agent, and the unlearn request for the respective agent is transmitted towards the non-RT RIC (403). The non-RT RIC (403) then removes the contributions from the agent identified in the unlearn request. At steps 427, performance is evaluated and optimized. Performance may be monitored and evaluated, the global model may be versioned, and updates to the global model are transmitted to the Near-RT RIC (407).

[0068] In some embodiments, the proposed solution would require an interface between the distributed components to communicate within each other. The communication link and protocol should support any distributed node to be able to send an unlearn request to the centralized unit at any time.

[0069] FIG.5 illustrates a method, according to some embodiments. In some embodiments, method 500 is a computer-implemented method for machine unlearning in federated learning. In some embodiments, method 500 is performed by a server mode.

[0070] Step s502 of the method includes obtaining a global model, wherein the global model is constructed using contribution weights and disentangled model parameters associated with a plurality of agents.

[0071] Step s504 of the method includes obtaining an unlearn request, the unlearn request indicating a first agent of the plurality of agents.

[0072] Step s506 of the method includes identifying a first contribution weight for a first set of disentangled model parameters associated with the first agent.

[0073] Step s508 of the method includes reducing the identified first contribution weight.

[0074] Step s510 of the method includes updating the global model based on the reduced first contribution.

[0075] Step s512 of the method includes generating a message comprising the updated global model.

[0076] Step s514 of the method includes transmitting the message towards a second agent.

[0077] FIG.6 is a block diagram of a node or apparatus 600 according to some embodiments. As shown in FIG. 6, the node may comprise: processing circuitry (PC) 602, which may include one or more processors (P) 655 (e.g., one or more general purpose microprocessors and / or one or more other processors, such as an application specific integrated circuit (ASIC), field- programmable gate arrays (FPGAs), and the like); communication circuitry 648, comprising a transmitter (Tx) 645 and a receiver (Rx) 647 for enabling the node to transmit data and receive data (e.g., wirelessly transmit / receive data) over network 610; and a local storage unit (a.k.a., “data storage system”) 608, which may include one or more non-volatile storage devices and / or one or more volatile storage devices. In embodiments where PC 602 includes a programmable processor, a computer program product (CPP) 641 may be provided. CPP 641 includes a computer readable medium (CRM) 642 storing a computer program (CP) 643 comprising computer readable instructions (CRI) 644. CRM 642 may be a non-transitory computer readable medium, such as, magnetic media (e.g., a hard disk), optical media, memory devices (e.g., random access memory, flash memory), and the like. In some embodiments, the CRI 644 of computer program 643 is configured such that when executed by PC 602, the CRI causes the apparatus to perform steps described herein (e.g., steps described herein with reference to the flow charts). In other embodiments, the apparatus may be configured to perform steps described herein without the need for code. That is, for example, PC 602 may consist merely of one or more ASICs. Hence, the features of the embodiments described herein may be implemented in hardware and / or software.

[0078] Set forth below is a concise summary of embodiments of the present disclosure. 1. A computer-implemented method for machine unlearning in federated learning, the method comprising: obtaining a global model, wherein the global model is constructed using contribution weights and disentangled model parameters associated with a plurality of agents;obtaining an unlearn request, the unlearn request indicating a first agent of the plurality of agents; identifying a first contribution weight for a first set of disentangled model parameters associated with the first agent; reducing the identified first contribution weight; updating the global model based on the reduced first contribution weight; generating a message comprising the updated global model; and transmitting the message towards a second agent. 2. The method of embodiment 1, further comprising: receiving an acknowledgement transmitted by the second agent, wherein the acknowledgement indicates that a local model of the second agent has been updated based on the updated global model. 3. The method of embodiment 1 or 2, wherein the message comprising the updated global model further comprises a signature of the updated global model and an indication for the second agent to update the local model of the second agent, and the signature of the updated global model is for validating the updated global model 4. The method of any one of embodiments 1-3, further comprising: obtaining, for each agent of the plurality of agents, a local model comprising a plurality of local model parameters trained on local data available to the respective agent. 5. The method of embodiment 4, further comprising: encoding the local model parameters into a column matrix of latent contribution vectors, wherein each row or column of the matrix corresponds to an agent or a group of agents of the plurality of agents. 6. The method of embodiment 5, further comprising:obtaining a latent support set comprising a second group of latent variables, wherein the latent support set is used to project a latent contribution vector into a space where it can be compared against parameters of the global model. 7. The method of embodiment 5 or 6, further comprising: processing each latent contribution vector through a disentanglement functionality; and outputting, as a result of the processing, a disentangled latent contribution vector that is disentangled from other latent contribution vectors. 8. The method of embodiment 7, wherein the disentanglement functionality comprises enforcing orthogonality across latent contributions. 9. The method of any one of embodiments 6-8, further comprising: calculating contribution weights for each agent of the plurality of agents by summing all elements of a disentangled latent contribution vector for each respective agent. 10. The method of embodiment 9, further comprising: decoding each disentangled latent contribution vector; and outputting, as a result of the decoding, disentangled model parameters. 11. The method of embodiment 10, further comprising: constructing the global model by averaging the disentangled model parameters weighted by the contribution weights. 12. The method of any one of embodiments 1-11, wherein the unlearn request is based on at least one of: a cyber attack against the first agent, a security breach at the first agent; poor performance of a model due to poor input from the first agent, detected erroneous model parameters at the first agent, a data privacy request relating to training data at the first agent, orbiased or untrustworthy actions output at the first agent. 13. The method of any one of embodiments

[0011] -12, wherein the reducing comprises reducing the identified first contribution weight to zero. 14. The method of any one of embodiments 1-13, wherein the updating the global model based on the reduced first contribution weight comprises at least one of: excluding original input from a local model of the first agent without the need to retrain the global model or reducing an influence of the original input from the local model on the global model. 15. The method of any one of embodiments 1-14, wherein the unlearn request indicates a group of agents of the plurality of agents, wherein the group of agents comprises the first agent, and wherein the first contribution weight is associated with the group of agents. 16. A computing device comprising processing circuitry and a memory coupled to the processing circuitry, wherein the device is adapted to: obtain a global model, wherein the global model is constructed using contribution weights and disentangled model parameters associated with a plurality of agents; obtain an unlearn request, the unlearn request indicating a first agent of the plurality of agents; identify a first contribution weight for a first set of disentangled model parameters associated with the first agent; reduce the identified first contribution weight; update the global model based on the reduced first contribution weight; generate a message comprising the updated global model; and transmit the message towards a second agent. 17. The computing device of embodiment 16, wherein the device is adapted to perform any one of the methods 2-15.18. A computer program comprising instructions which when executed by processing circuity of a computing device causes the device to perform the method of any one of methods 1- 15. 19. A carrier containing the computer program of embodiment 17, wherein the carrier is one of an electronic signal, an optical signal, a radio signal, and a computer readable storage medium. 20. An apparatus, the apparatus comprising: communication circuitry; and processing circuitry, wherein the apparatus is configured to: obtain a global model, wherein the global model is constructed using contribution weights and disentangled model parameters associated with a plurality of agents; obtain an unlearn request, the unlearn request indicating a first agent of the plurality of agents; identify a first contribution weight for a first set of disentangled model parameters associated with the first agent; reduce the identified first contribution weight; update the global model based on the reduced first contribution weight; generate a message comprising the updated global model; and transmit the message towards a second agent. 21. An apparatus, the apparatus adapted to perform the method of any one of the methods of embodiments 1-15.

[0079] While various embodiments are described herein, it should be understood that they have been presented by way of example only, and not limitation. Thus, the breadth and scope of this disclosure should not be limited by any of the above described embodiments. Moreover, any combination of the above-described elements in all possible variations thereof is encompassed by the disclosure unless otherwise indicated herein or otherwise clearly contradicted by context.

[0080] Additionally, while the processes described above and illustrated in the drawings are shown as a sequence of steps, this was done solely for the sake of illustration. Accordingly, it is contemplated that some steps may be added, some steps may be omitted, the order of the steps may be re-arranged, and some steps may be performed in parallel.

[0081] REFERENCES

[0082] [1] Halimi, A., Kadhe, S., Rawat, A., & Baracaldo, N. (2022). Federated Unlearning: How to Efficiently Erase a Client in FL? ArXiv, abs / 2207.05521.

[0083] [2] Wu, C., Zhu, S., & Mitra, P. (2022). Federated Unlearning with Knowledge Distillation. ArXiv, abs / 2201.09441.

[0084] [3] Liu, G., Ma, X., Yang, Y., Wang, C., & Liu, J. (2021). FedEraser: Enabling Efficient Client-Level Data Removal from Federated Learning Models. 2021 IEEE / ACM 29th International Symposium on Quality of Service (IWQOS), 1-10.

Claims

CLAIMS 1. A computer-implemented method for machine unlearning in federated learning, the method comprising: obtaining a global model, wherein the global model is constructed using contribution weights and disentangled model parameters associated with a plurality of agents; obtaining an unlearn request, the unlearn request indicating a first agent of the plurality of agents; identifying a first contribution weight for a first set of disentangled model parameters associated with the first agent; reducing the identified first contribution weight; updating the global model based on the reduced first contribution weight; generating a message comprising the updated global model; and transmitting the message towards a second agent.

2. The method of claim 1, further comprising: receiving an acknowledgement transmitted by the second agent, wherein the acknowledgement indicates that a local model of the second agent has been updated based on the updated global model.

3. The method of claim 1 or 2, wherein the message comprising the updated global model further comprises a signature of the updated global model and an indication for the second agent to update the local model of the second agent, and the signature of the updated global model is for validating the updated global model.

4. The method of any one of claims 1-3, further comprising: obtaining, for each agent of the plurality of agents, a local model comprising a plurality of local model parameters trained on local data available to the respective agent.

5. The method of claim 4, further comprising:encoding the local model parameters into a matrix of latent contribution vectors, wherein each row or column of the matrix corresponds to an agent or a group of agents of the plurality of agents.

6. The method of claim 5, further comprising: obtaining a latent support set comprising a second group of latent variables, wherein the latent support set is used to project a latent contribution vector into a space where it can be compared against parameters of the global model.

7. The method of claim 5 or 6, further comprising: processing each latent contribution vector through a disentanglement functionality; and outputting, as a result of the processing, a disentangled latent contribution vector that is disentangled from other latent contribution vectors.

8. The method of claim 7, wherein the disentanglement functionality comprises enforcing orthogonality across latent contributions.

9. The method of any one of claims 6-8, further comprising: calculating contribution weights for each agent of the plurality of agents by summing all elements of a disentangled latent contribution vector for each respective agent.

10. The method of claim 9, further comprising: decoding each disentangled latent contribution vector; and outputting, as a result of the decoding, disentangled model parameters.

11. The method of claim 10, further comprising: constructing the global model by averaging the disentangled model parameters weighted by the contribution weights.

12. The method of any one of claims 1-11, wherein the unlearn request is based on at least one of:a cyber attack against the first agent, a security breach at the first agent; poor performance of a model due to poor input from the first agent, detected erroneous model parameters at the first agent, a data privacy request relating to training data at the first agent, or biased or untrustworthy actions output at the first agent.

13. The method of any one of claims 1-12, wherein the reducing comprises reducing the identified first contribution weight to zero.

14. The method of any one of claims 1-13, wherein the updating the global model based on the reduced first contribution weight comprises at least one of: excluding original input from a local model of the first agent without the need to retrain the global model or reducing an influence of the original input from the local model on the global model.

15. The method of any one of claims 1-14, wherein the unlearn request indicates a group of agents of the plurality of agents, wherein the group of agents comprises the first agent, and wherein the first contribution weight is associated with the group of agents.

16. A computing device comprising processing circuitry and a memory coupled to the processing circuitry, wherein the device is adapted to: obtain a global model, wherein the global model is constructed using contribution weights and disentangled model parameters associated with a plurality of agents; obtain an unlearn request, the unlearn request indicating a first agent of the plurality of agents; identify a first contribution weight for a first set of disentangled model parameters associated with the first agent; reduce the identified first contribution weight; update the global model based on the reduced first contribution weight; generate a message comprising the updated global model; andtransmit the message towards a second agent.

17. The computing device of claim 16, wherein the device is adapted to perform any one of the method of any one of claims 2-15.

18. A computer program comprising instructions which when executed by processing circuity of a computing device causes the device to perform the method of any one of claims 1- 15.

19. A carrier containing the computer program of claim 18, wherein the carrier is one of an electronic signal, an optical signal, a radio signal, and a computer readable storage medium.

20. An apparatus, the apparatus comprising: communication circuitry; and processing circuitry, wherein the apparatus is configured to: obtain a global model, wherein the global model is constructed using contribution weights and disentangled model parameters associated with a plurality of agents; obtain an unlearn request, the unlearn request indicating a first agent of the plurality of agents; identify a first contribution weight for a first set of disentangled model parameters associated with the first agent; reduce the identified first contribution weight; update the global model based on the reduced first contribution weight; generate a message comprising the updated global model; and transmit the message towards a second agent.

21. An apparatus, the apparatus adapted to perform the method of any one of claims 1-15.

Images