Confidential distributed machine learning

By using a secure multi-party computing and trusted execution environment in federated learning, ensuring that the secret shared representation of machine learning models is securely aggregated on the aggregator cluster, solving the problem that local training data confidentiality in traditional federated learning is difficult to ensure, and achieving more High security and confidentiality protection.

CN119026669APending Publication Date: 2024-11-26ROBERT BOSCH GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410649544.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2023-05-24
Filing Date
2024-05-23
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Traditional federated learning is difficult to fully guarantee the confidentiality of local training data, especially the possibility of malicious aggregators exporting training data from machine learning model updates, and global machine learning models are vulnerable to inference attacks.

Method used

Upload secret shared representations of machine learning models to the aggregator cluster through secure multi-party computing (MPC) and trusted execution environment (TEE), and secure aggregation of machine learning models on the aggregator cluster ensures only identifiers rather than The model itself is transmitted and processed to prevent malicious attacks.

Benefits of technology

Significantly improves the confidentiality of machine learning model updates, prevents confidentiality attacks, enhances security, and protects training clients from malicious aggregators and orchestrators.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026669A_ABST
    Figure CN119026669A_ABST
Patent Text Reader

Abstract

The invention relates to confidential distributed machine learning. A general aspect of the present disclosure relates to a method, the methods include a computer-implemented method for federated learning for an owner of a machine learning model, a computer-implemented method for federated learning for an orchestrator, and a computer-implemented method for federated learning for an orchestrator. A computer-implemented method for federated learning for a training client and / or a computer-implemented method for federated learning for an aggregator.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to confidential distributed machine learning. Background Art

[0002] Federated (machine) learning can be considered a subgenus of distributed (machine) learning. In federated machine learning (see, for example, Priyanka Mary Mammen. 2021. Federated Learning: Opportunities and Challenges. In Proceedings of ACM Conference (Conference'17). ACM, New York, NY, USA, 5 pages; http: / / arxiv.org / pdf / 2010.0528.pdf ), multiple devices learn a machine learning model, such as an artificial neural network, together, for example under the supervision of a central server, without sharing their private training data. Federated (machine) learning can therefore be used in particular where private training data cannot be shared for data protection reasons (e.g. in the health sector, the financial sector, etc.).

[0003] In federated learning, local machine learning model updates are generated on participating devices through local training processes, and the local machine learning model updates are aggregated (i.e., combined) into a trained global machine learning model. Algorithms for secure aggregation are, for example, from "SAFELearn: Secure Aggregation for private FEderated Learning", Hossein Fereidooni et al., Cryptology ePrint Archive, Paper 2021 / 386; https: / / eprint.iacr.org / 2021 / 386 Known.

[0004] A Trusted Execution Environment (TEE) provides a secure or trusted runtime environment for applications. A known example of a TEE is Intel Software Guard Extensions (Intel SGX), see, for example, https: / / de.wikipedia.org / w / index.php?title=Software_Guard_Extensions&oldid=232528710. The concept of remote attestation is known in the context of trusted computing. Thus, for example, changes at a user's computer can be identified by an authorized party, see, for example, https: / / en.wikipedia.org / w / index.php?title=Trusted_Computing&oldid=1151565594#Remote_attestation.

[0005] (Secure) multi-party computation (MPC) - in English, for example: (Secure) multi-party computation (MPC) - is also a branch of cryptography, the goal of which is to develop methods with which parties can jointly compute functions via their input parameters, wherein these input parameters are kept secret, see, for example, https: / / en.wikipedia.org / w / index.php?title=Secure_multi-party_computation&oldid=1148234769. Cloud-native (secure) multi-party computation can be implemented, for example, by the Carbyne Stack, see, for example, https: / / carbynestack.io.

[0006] This disclosure proposes measures to improve confidentiality in federated learning. Summary of the invention

[0007] A first general aspect of the present disclosure relates to a computer-implemented method for federated learning for owners of machine learning models. The method includes uploading a first secret share of a first multi-party computation (MPC) representation (CS(G_i)) of a machine learning model (G_i) to a cluster of an aggregator, wherein the machine learning model (G_i) on the cluster of the aggregator is federated learning. _ i), sends a first identifier (ID(CS(G_i))) of a first MPC representation (CS(G_i)) of the first MPC representation (CS(G_i)) to the owner. The method further comprises sending a first trigger signal comprising the first identifier (ID(CS(G_i))) to the orchestrator.

[0008] A second general aspect of the present disclosure relates to a computer-implemented method for performing federated learning for an orchestrator. The method includes sending a second trigger signal including a first identifier (ID(CS(G_i))) of a first MPC representation (CS(G_i)) of a machine learning model (G_i) on a cluster of an aggregator to a plurality of (C) training clients (c_j) for training iterations if a first trigger signal including a first identifier (ID(CS(G_i))) has been received. The method further includes sending a fourth trigger signal including a second identifier (ID(CS(L_i_j))) of a second MPC representation (CS(L_i_j)) of a local machine learning model update (L_i_j) on a cluster of an aggregator to an aggregator if at least one third trigger signal including a second identifier (ID(CS(L_i_j))) of a second MPC representation (CS(L_i_j)) of a local machine learning model update (L_i_j)) on a cluster of an aggregator has been received.

[0009] A third general aspect of the present disclosure relates to a computer-implemented method for performing federated learning for a training client (c_j). The method includes downloading a first MPC representation (CS(G_i)) of a machine learning model (G_i) from the cluster of the aggregator based on the first identifier (ID(CS(G_i))) if a second trigger signal including a first identifier (ID(CS(G_i))) has been received and the identity and authority of the training client (c_j) have been verified on the cluster of the aggregator. The method further includes converting the first MPC representation (CS(G_i)) of the machine learning model into a local machine learning model according to a predetermined MPC protocol. The method further includes training a local machine learning model based on local training data of the training client, wherein a local machine learning model update (L_i_j) is generated. The method further includes converting the local machine learning model update (L_i_j) into a second MPC representation (CS(L_i_j)) according to a predetermined MPC protocol. The method further includes uploading a second secret share of a second MPC representation (CS(L_i_j)) of the local machine learning model update (L_i_j) to the cluster of the aggregator, wherein a second identifier (ID(CS(L_i_j))) of the second MPC representation (CS(L_i_j)) of the local machine learning model update (L_i_j) on the cluster of the aggregator is sent to the training client (c_j). The method includes sending a third trigger signal including the second identifier (ID(CS(L_i_j))) to the orchestrator.

[0010] A fourth general aspect of the present disclosure relates to a computer-implemented method for performing federated learning for an aggregator. The method includes sending a first identifier (ID (CS (G_i))) of a first MPC representation (CS (G_i)) of a machine learning model (G_i) on the cluster of the aggregator to an owner of the machine learning model (G_i) if a first secret share of the first MPC representation (CS (G_i)) is uploaded to the cluster of the aggregator. The method further includes sending a second identifier (ID (CS (L_i_j))) of a second MPC representation (CS (L_i_j)) of a local machine learning model update (L_i_j) on the cluster of the aggregator to a training client (c_j) if a second secret share of the second MPC representation (CS (L_i_j)) is uploaded to the cluster of the aggregator. The method further includes, if a fourth trigger signal is received, securely aggregating the local machine learning model update (L_i_j) with at least one other local machine learning model update based on a predetermined MPC circuit, wherein a third secret sharing of a third MPC representation (CS(G_(i+1))) of the aggregated machine learning model (G_(i+1)) on the cluster of the aggregator and a third identifier (ID(CS(G_(i+1))))) of the third MPC representation (CS(G_(i+1))) are obtained. The method further includes sending the third identifier (ID(CS(G_(i+1)))) to the orchestrator.

[0011] A fifth general aspect of the present disclosure relates to a method comprising the method according to the first general aspect, the method according to the second general aspect, the method according to the third general aspect and / or the method according to the fourth general aspect.

[0012] A sixth general aspect of the present disclosure relates to a device designed to perform a method according to any one of the preceding general aspects.

[0013] A seventh general aspect of the present disclosure relates to a computer program designed to execute the method according to any one of the preceding general aspects.

[0014] An eighth general aspect of the present disclosure relates to a data carrier or a signal comprising / encoding a computer program according to the seventh general aspect.

[0015] In the case of federated (machine) learning, multiple (usually a large number of) devices (referred to below as training clients) train a (global) machine learning model, such as an (artificial) neural network, in common. The training clients do not necessarily have to suffer the same responsibilities in this case. Instead, the training clients may be located in different locations and / or suffer different responsibilities. An exemplary but non-exclusive scenario includes one (local) training client per clinic, each of which has (local) training data (e.g., diagnostic photos, patient data, etc.), which the clinic cannot share with other clinics or instances, but is still ready to contribute based on these corresponding (local) training data for training the (global) machine learning model. In general, this can be achieved in such a way that each (local) training client obtains a (global) machine learning model and generates a (local) machine learning model update based on its (local) training data, wherein, for example, the (global) machine learning model is trained on the training client using the corresponding local training data. Then, all (local) machine learning model updates are aggregated by a central instance (which can be referred to as an aggregator) so that a (global) machine learning model update and in particular a trained (global) machine learning model are obtained.

[0016] Since participating training clients keep their corresponding (local) training data to themselves and do not transfer it to third parties, federated learning is generally considered to be a safer and more data protection-friendly solution for training machine learning models with sensitive data.

[0017] However, it has been repeatedly shown (see, for example, Priyanka Mary Mammen. 2021. Federated Learning: Opportunities and Challenges. In Proceedings of ACM Conference (Conference'17). ACM, New York, NY, USA, 5 pages; https: / / arxiv.org / odf / 2101.0542.pdf and Benmalek et al. Security of Federated Learning: Attacks, Defensive Mechanisms, and Challenges. Revue des Sciences et Technologies de 1'Information-Série RIA: Revue d'Intelligence Artificielle, 2022, 36(1), pp.49-59, 10.18280 / ria.360106, hal-0362040, https: / / hal.science / hal-03620400 / document), traditional federated learning is far from being able to fully guarantee the confidentiality of local training data. In particular, training data may be derived from machine learning model updates by malicious aggregators. In addition, global machine learning models may be vulnerable to inference attacks that aim to derive training data from the output data of (locally) trained machine learning models.

[0018] The confidentiality attack may include a membership inference attack that aims to determine whether a particular training data was used by a training client during training. Alternatively or additionally, the confidentiality attack may include an attribute inference attack that aims to derive meta-features of training data of other training clients. Alternatively or additionally, the confidentiality attack may include a reconstruction attack that aims to reconstruct the training data and / or labels associated therewith that have been used during training.

[0019] Due to the methods of the present disclosure, such confidentiality attacks on machine learning model updates can be at least made difficult or completely prevented. In addition to confidentiality, security can also be improved due to the methods of the present disclosure. In particular, these methods protect the training client from confidentiality attacks emanating from malicious aggregators and / or malicious orchestrators designed to coordinate federated learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1a Schematically illustrates a computer-implemented method for federated learning with owners of machine learning models.

[0021] Figure 1b Schematically illustrating a computer-implemented method for federated learning for an orchestrator.

[0022] Figure 1c Schematically illustrates a computer-implemented method for federated learning for a training client.

[0023] Figure 1d Schematic illustration of a computer-implemented method for federated learning for an aggregator.

[0024] Figure 2 The schematic diagram includes at least one Figure 1a -d method as shown in the method.

[0025] Figure 3 Schematic illustration of a machine learning model's owner, orchestrator, training client, aggregator, and optionally tester. DETAILED DESCRIPTION

[0026] The computer-implemented method proposed in the present disclosure is performed in the interaction of the following entities: Figure 3: The owner of the machine learning model, multiple (usually large) training clients, orchestrators, and aggregators. Additionally, testers may also be used.

[0027] The owner 10 of a (global) machine learning model has rights to the (global) machine learning model, not only in the initial state but also in the trained state. In particular, the owner can choose the machine learning model, i.e. its architecture (e.g. neurons, layers, number of neurons per layer, etc.). The owner is also a customer of the federated trained machine learning model.

[0028] A training client c_j (i.e. j=1 to m) among the plurality / large number C training clients has (local) training data that should at least not be shared with a third party or may not be shared with a third party. A training client may be a device (e.g. a computer) within an organizational unit (e.g. a clinic) that has (local) training data.

[0029] The orchestrator 20 coordinates federated learning by triggering actions on training clients c_j. Coordination includes, for example, selecting the training client that should be used in the next training iteration, providing references to the initial and updated (i.e., trained) machine learning models, and / or evaluating the training progress in the training iteration. In addition, the orchestrator delegates the aggregation 40 of the machine learning model updates of the training clients to the aggregator. Since communication is done through identifiers, the orchestrator never sees the machine learning model updates or the machine learning models themselves.

[0030] The aggregator 40 comprises a (secure) multi-party computation (MPC) cluster, i.e. a large number of devices (e.g. computers) designed in particular to operate in the context of MPC. This means that each cluster member of the MPC cluster participates in the common computation of the cluster. The input of each individual cluster member comprises a secret component of a secret share of the total input, so that no individual cluster member learns the content of the total input, the content of its own secret component, or the content of the components of other cluster members. The aggregator (i.e. the MPC cluster) obtains machine learning model updates from training clients, computes aggregate model updates, and updates the global machine learning model accordingly. By using (secure) multi-party computation (MPC), the aggregator advantageously has no knowledge of the model updates and the global machine learning model.

[0031] The following conventions are used below:

[0032] G_i represents a machine learning model, more precisely the i-th version of the machine learning model (e.g., i=0 to n-1). G_0 may, for example, be an initial machine learning model for executing the disclosed method. For example, G_0 may be an untrained machine learning model. Alternatively, G_0 may be a machine learning model that has been previously trained by the disclosed method or otherwise. For example, G_i may be a data structure that defines model parameters of a machine learning model.

[0033] For example, CS(x) where x=G_i or L_i_j represents a multi-party computation (MPC) representation of x. Such a representation is reversible, i.e., x can be converted into an MPC representation CS(x) according to a predetermined MPC protocol, and the MPC representation CS(x) can also be converted into x according to a predetermined MPC protocol. For example, CS(x) can be an MP-SPDZ / Carbyne Stack representation.

[0034] ID(y) is the identifier of the secret share y stored on the aggregator's MPC cluster. For this purpose, for example in the case of CarbyneStack, the Amphora service of the cluster members can be used.

[0035] L_i_j represents the machine learning model update (locally) generated by the training client c_j in the i-th training iteration (e.g., i=0 to n-1). The machine learning model update can be, for example, a machine learning model updated by training.

[0036] First, a computer-implemented method 100 for performing federated learning for an owner 10 of a machine learning model is disclosed. Figure 1a It is schematically shown in FIG.

[0037] The method 100 includes uploading 130 a first secret share of a first multi-party computation (MPC) representation CS(G_i) of a (e.g., initial) machine learning model G_i (i.e., e.g., G_0) to a cluster of an aggregator 40, wherein a first identifier ID (CS(G_i)) of the first MPC representation CS(G_i) of the machine learning model G_i on the cluster of the aggregator 40 is sent to the owner 10. The machine learning model uploaded 130 may be an initial, i.e., untrained machine learning model (e.g., G_0). Alternatively, the machine learning model uploaded 130 may be a machine learning model G_i that has been at least partially trained.

[0038] The method 100 further comprises sending 131 a first trigger signal comprising a first identifier ID(CS(G_i)) to the orchestrator 20. The first trigger signal may, for example, consist of only the first identifier ID(CS(G_i)). The first trigger signal may prompt the orchestrator 20 to perform federated learning on the training client. Providing the machine learning model G_i as a secret share and otherwise only providing its identifier may have the advantage that the orchestrator 20 does not obtain the machine learning model itself. The confidentiality of the machine learning model may thereby be better ensured.

[0039] The method 100 may further include converting 110 the (e.g. initial) machine learning model G_i into a first MPC representation CS(G_i) according to a predetermined MPC protocol.

[0040] The method 100 may further include downloading 140 a third MPC representation CS(G_(i+1)) of the aggregated machine learning model G_(i+1) from the cluster of the aggregator 40 based on the third MPC representation CS(G_(i+1)) of the aggregated machine learning model G_(i+1) on the cluster of the aggregator 40 if the third identifier ID(CS(G_(i+1))) has been received. Then, the method 100 may further include converting 141 the third MPC representation CS(G_(i+1)) of the aggregated machine learning model G_(i+1) into a global machine learning model according to a predetermined MPC protocol. Through steps 140 and 141, the owner of the machine learning model can call the machine learning model in a trained state. In the case of, for example, n training iterations, first CS(G_(n-1)) may also be downloaded as a third MPC representation.

[0041] In addition, a computer-implemented method 200 for performing federated learning for an orchestrator 20 is disclosed. Figure 1b It is schematically shown in FIG.

[0042] The method 200 comprises: if a first trigger signal comprising a first identifier ID (CS(G_i)) has been received, a second trigger signal is sent 230 to a plurality of C (typically even a large number) of training clients c_j for training iterations, wherein the second trigger signal comprises a first identifier ID (CS(G_i)) of a first MPC representation CS(G_i) of a machine learning model G_i on a cluster of the aggregator 40. The second trigger signal can be used to prompt the training clients c_j to perform training iterations respectively. The second trigger signal can, for example, consist only of the first identifier ID (CS(G_i)). In general, not only the orchestrator but also each training client c_j knows only the first identifier ID (CS(G_i)), but not the machine learning model G_i (or its MPC representation CS(G_i)). The latter is stored on the cluster of the aggregator 40 as a secret share in a completely trustworthy manner.

[0043] The method 200 may include selecting 210 a plurality / large number C of training clients c_j for training iterations according to a predetermined selection strategy.

[0044] Furthermore, the method 200 comprises sending 231 a fourth trigger signal to the aggregator 40, the fourth trigger signal comprising a second identifier ID(CS(L_i_j)) of a second MPC representation CS(L_i_j) of the local machine learning model update L_i_j on the cluster of the aggregator 40, if at least one third trigger signal has been received, the third trigger signal comprising a second identifier ID(CS(L_i_j)) of a second MPC representation CS(L_i_j) of the local machine learning model update L_i_j on the cluster of the aggregator 40. The third trigger signal may, for example, consist of only the second identifier ID(CS(L_i_j)). Similarly, the fourth trigger signal may, for example, consist of only the second identifier ID(CS(L_i_j)).

[0045] If another third trigger signal for training iteration has been received by at least one other training client c_j (for different j) among multiple / large numbers of C training clients c_j, and the other third trigger signal includes another second identifier ID (CS((L_i_j)) of another second MPC representation CS(L_i_j) of another local machine learning model update L_i_j on the cluster of the aggregator 40, a fourth trigger signal can be sent 231, wherein the fourth trigger signal includes the other second identifier ID (CS(L_i_j)) (for different j).

[0046] If each of the multiple / large number of C training clients c_j has received a corresponding third trigger signal for a training iteration, wherein the corresponding third trigger signal includes a corresponding second identifier ID (CS((L_i_j)) of the corresponding second MPC representation CS(L_i_j) of the corresponding local machine learning model update L_i_j on the cluster of the aggregator 40, then in particular a fourth trigger signal can be sent 231, wherein the fourth trigger signal includes each second identifier ID (CS(L_i_j)) (j=1 to m).

[0047] Method 200 may further include checking 240 whether another training iteration should be performed for the aggregated machine learning model G_(i+1) if a third identifier ID(CS(G_(i+1))) of a third MPC representation CS(G_(i+1)) of the aggregated machine learning model G_(i+1) on the cluster of the aggregator 40 has been received.

[0048] Then, the method 200 may further include: if another training iteration should not be performed, sending 241 a third identifier ID(CS(G_(i+1))) to the owner 10 of the machine learning model. Otherwise, the method may be repeated for each further training iteration (e.g., n training iterations), wherein, for example, ID(CS(G_(n-1))) is returned to the owner 10 of the machine learning model as a third identifier.

[0049] Methods 100, 200, 300, 400, 500 have protected the machine learning model from malicious training clients. In addition, the method also protects the machine learning model from being disclosed to the orchestrator 20. However, there is still a possible danger to the extent that the orchestrator 20 may negatively affect the quality of the (trained) machine learning model due to the maliciously distorted selection of the training client. This can be prevented as follows:

[0050] Method 200 may be executed within a trusted execution environment (TEE). This increases security, in particular because orchestrator 20 is prevented from maliciously deviating from method 200 and / or the type of settings. Here, the owner 10 of the machine learning model may trust orchestrator 20 to adhere to a predetermined selection policy, for example, through remote attestation. Thus, malicious distorted selections by training clients may be prevented.

[0051] Furthermore, in the method 200, the log entries can be recorded in a completely reliable manner on an external memory, optionally on a distributed ledger. This can also increase security, since the process can also be understood afterwards. For example, completely reliable recording can be achieved by encrypting the log entries in an asymmetrical manner using the public key of the owner 10 of the machine learning model. Alternatively, the log entries can be encrypted using a private symmetric key previously agreed upon by the owner 10 of the machine learning model and the orchestrator 20.

[0052] In addition, in method 200, the orchestrator 20 may prompt 250 at least one training client c_j to download the aggregated machine learning model G_(i+1) and evaluate the quality of the aggregated machine learning model G_(i+1) based on the local test data of the training client (according to predetermined local test criteria), wherein at least one second test result is obtained and sent to the orchestrator 20.

[0053] The method 200 may then include receiving 251 at least a second test result.

[0054] Then, method 200 may include evaluating 252 at least a second test result and a first test result obtained by a tester 50, wherein the tester 50 is designed to evaluate the quality of the aggregate machine learning model G_(i+1) based on local test data of the tester 50 (also according to predetermined test criteria), wherein the evaluation is based on MPC or homomorphic encryption, wherein an evaluation result is obtained.

[0055] The method 200 may then include performing 253 one or more predetermined actions according to the evaluation results. For example, the predetermined action may be to omit other training iterations because the evaluation results are already satisfactory. Alternatively, the predetermined action may also be, for example, to switch to other predetermined selection strategies for training clients.

[0056] In addition, a computer-implemented method 300 for performing federated learning for a training client c_j is disclosed. Figure 1c Shown schematically.

[0057] The method 300 comprises: if a second trigger signal comprising a first identifier ID(CS(G_i)) has been received and the identity and authority of the training client c_j has been verified on the cluster of the aggregator 40, downloading 320 a first MPC representation CS(G_i) of the machine learning model G_i from the cluster of the aggregator 40 based on the first identifier ID(CS(G_i)). Verifying the identity and authority of the training client may be based, for example, on a predetermined protocol for authentication and / or authorization.

[0058] The method 300 then comprises converting 321 the first MPC representation CS(G_i) of the machine learning model into a local machine learning model according to a predetermined MPC protocol (the same MPC protocol as in the method 100). The local machine learning model may be, for example, the machine learning model G_i and in particular the initial machine learning model G_0. Alternatively, the local machine learning model may be another representation of the machine learning model G_i (in particular G_0) designed for training on the training client c_j.

[0059] The method 300 further comprises training 330 a local machine learning model based on local training data of the training client, wherein a local machine learning model update L_i_j is generated. The local machine learning model update L_i_j may be, for example, a trained local machine learning model.

[0060] The method 300 further includes converting 340 the local machine learning model update L_i_j into a second MPC representation CS(L_i_j) according to a predetermined MPC protocol.

[0061] Method 300 further includes uploading 342 a second secret share of a second MPC representation CS(L_i_j) of the local machine learning model update L_i_j to the cluster of the aggregator 40, wherein a second identifier ID(CS(L_i_j)) of the second MPC representation CS(L_i_j) of the local machine learning model update L_i_j on the cluster of the aggregator 40 is sent to the training client c_j.

[0062] The method 300 then comprises sending 343 a third trigger signal comprising the second identifier ID(CS(L_i_j)) to the orchestrator 20 .

[0063] The method 300 can be executed in a trusted execution environment (TEE) (on at least one training client c_j or on each training client c_j of a plurality of C training clients). This ensures that one or more training clients c_j—preferably all training clients—cannot (maliciously) deviate from the method 300. This can prevent, for example, the machine learning model and in particular its misuse.

[0064] Also disclosed is a computer-implemented method 400 for performing federated learning on an aggregator 40. Figure 1d It is schematically shown in FIG.

[0065] Method 400 includes sending 410 a first identifier ID(CS(G_i)) of the first MPC representation CS(G_i) of the machine learning model G_i on the cluster of aggregator 40a to the owner 10 of the machine learning model G_i if the first secret share of the first MPC representation CS(G_i) is uploaded to the cluster of aggregator 40.

[0066] The method 400 further includes updating the local machine learning model on the cluster of aggregators 40 with a second identifier ID(CS(L_i_j)) of the second MPC representation CS(L_i_j) of L_i_j if the second secret share of the second MPC representation CS(L_i_j) is uploaded to the cluster of aggregators 40

[0067] Send 420 to the training client c_j.

[0068] The method 400 further includes securely aggregating 431 the local machine learning model update L_i_j with at least one other local machine learning model update based on a predetermined MPC circuit if a fourth trigger signal is received, wherein a third secret sharing of a third MPC representation CS(G_(i+1)) of the aggregated machine learning model G_(i+1) on the cluster of the aggregator 40 and a third identifier ID(CS(G_(i+1))) of the third MPC representation CS(G_(i+1)) are obtained. The secure aggregation 431 can be implemented, for example, by or based on an algorithm from SAFELeam, see above. For example, the aggregated machine learning model G_(i+1) can be based on the weighting of the local machine learning model updates L_i_j (for different j).

[0069] The method 400 then includes sending 432 the third identifier ID(CS(G_(i+1))) to the orchestrator 20 .

[0070] Alternatively, after n training iterations, the aggregate machine learning model CS(G_(n-1)) can be aggregated and the corresponding identifier ID(CS(G_(n-1))) can be sent 432 to the orchestrator 20.

[0071] Method 400 may further include sending 440 a third identifier ID(CS(G_(i+1))) of a third MPC representation CS(G_(i+1)) to a tester 50, wherein the tester is designed to download the aggregated machine learning model G-(i+1) and evaluate the quality of the aggregated machine learning model G-(i+1) based on local test data of the tester 50 (e.g., according to predetermined test criteria), wherein the evaluation is based on MPC or homomorphic encryption, wherein a first test result is obtained, and the test result is sent to the aggregator 40.

[0072] The method 400 may then include receiving 441 a first test result.

[0073] Method 400 may then include sending 442 the first test result to orchestrator 20. Based on the first test result, one or more predetermined actions may be performed by orchestrator 20 in method 100, see above.

[0074] Alternatively, the evaluation can be based on the machine learning model CS(G_(n-1)) that converges after n training iterations.

[0075] In addition, public Figure 2 One or more combined methods 500 schematically illustrated in FIG.

[0076] Method 500 may include computer-implemented method 100 for performing federated learning for owner 10 of machine learning model. Alternatively or additionally, method 500 may include computer-implemented method 200 for performing federated learning for orchestrator 20. Alternatively or additionally, method 500 may include computer-implemented method 300 for performing federated learning for training client c_j. Alternatively or additionally, method 500 may include computer-implemented method 400 for performing federated learning for aggregator 40.

[0077] The predetermined MPC protocol may be based on fixed point numbers, floating point numbers and / or integers (e.g., int8 quantization) in methods 100, 200, 300, 400, 500. The choice may be made based on the desired tradeoff between performance and accuracy of the machine learning model. Fixed point based variants are faster but may cause quantization errors, while floating point based variants require more computing power and furthermore use more bandwidth (in the network), but for this purpose enable a more accurate mapping between the original representation of the machine learning model and an MPC compatible representation.

[0078] Alternatively, methods 100, 200, 300, 400, 500 may be adapted so that the global machine learning model is accessible to orchestrator 20. In this variant, only local machine learning model updates are securely aggregated via MPC by aggregator 40. Here, the global machine learning model is exchanged in-band, i.e., over a regular communication channel, with the training client.

[0079] In addition, one or more devices are disclosed, each of which is designed to perform one or more methods 100, 200, 300, 400, 500. Each of these devices includes at least one computing unit (at least one processor) and a working memory (such as RAM), and may also have a non-volatile memory. The device may, for example, include a computing unit for the owner 10 of the machine learning model. For example, another device may be a computing unit for an orchestrator. For example, one or more other devices may each include at least one computing unit for training a client c_j. For example, another device may include at least one computing unit for an aggregator 40 (usually a large number of computing units for a cluster). The device may also include a computing unit for the owner 10 of the machine learning model, for the orchestrator 20, for each training client c_j, and / or for the aggregator 40. Such a device may, for example, be a computing unit in an associated network.

[0080] Furthermore, one or more computer programs are disclosed, which are respectively designed to perform one or more methods 100, 200, 300, 400, 500. For example, each of these computer programs can be in an interpretable or compiled form and can be (also partially) loaded into a working memory (e.g., RAM) of a device for execution, for example as a bit or byte sequence.

[0081] Furthermore, one or more data carriers or signals are disclosed, which respectively contain or encode the disclosed computer program. The data carrier may, for example, comprise one of a RAM, a ROM, an EPROM, a HDD, an SDD, ... on which / in which a signal is stored. The data carrier on which the computer program with the method 100, 200, 300, 400, 500 is stored may be a non-volatile memory of the device.

Claims

1. A computer-implemented method (100) for performing federated learning for an owner (10) of a machine learning model, comprising: Uploading (130) a first secret share of a first multi-party computation (MPC) representation (CS(G_i)) of a machine learning model (G_i) to a cluster of an aggregator (40), wherein a first identifier (ID(CS(G_i))) of the first MPC representation (CS(G_i)) of the machine learning model (G_i) on the cluster of the aggregator (40) is sent to the owner (10); A first trigger signal including the first identifier (ID(CS(G_i))) is sent (131) to the scheduler (20).

2. The method (100) according to claim 1, comprising: The machine learning model (G_i) is converted (110) into the first MPC representation (CS(G_i)) according to a predetermined MPC protocol.

3. The method (100) according to claim 1 or 2, comprising: downloading (140) a third MPC representation (CS(G_(i+1))) of the aggregated machine learning model (G_(i+1))) from the cluster of the aggregator (40) based on the third identifier (ID(CS(G_(i+1)))) of the aggregated machine learning model (G_(i+1))) on the cluster of the aggregator (40) if the third identifier (ID(CS(G_(i+1)))) has been received; The third MPC representation (CS(G_(i+1))) of the aggregate machine learning model (G_(i+1)) is converted (141) into a global machine learning model according to a predetermined MPC protocol.

4. A computer-implemented method (200) for performing federated learning for an orchestrator (20), comprising: If a first trigger signal including a first identifier (ID(CS(G_i))) has been received, a second trigger signal is sent (230) to a plurality (C) of training clients (c _ j) for a training iteration, the second trigger signal comprises a first identifier (ID(Cs(G_i))) of a first MPC representation (CS(G_i)) of a machine learning model (G_i) on a cluster of the aggregator (40); If at least one third trigger signal has been received, the third trigger signal comprising a second identifier (ID (CS (L_i_j))) of a second MPC representation (CS (L_i_j)) of a local machine learning model update (L_i_j) on a cluster of the aggregator (40)), a fourth trigger signal is sent (231) to the aggregator (40), the fourth trigger signal comprising a second identifier (ID (CS (L_i_j))) of a second MPC representation (CS (L_i_j)) of a local machine learning model update (L_i_j)) on a cluster of the aggregator (40)).

5. The method (200) according to claim 4, comprising: A plurality (C) of training clients (c_j) are selected (210) for training iterations according to a predetermined selection strategy.

6. The method (200) according to claim 4 or 5, wherein the fourth trigger signal is sent (231) if another third trigger signal has been received by at least one other training client (c_j) of the plurality (C) of training clients (cj) for a training iteration, the another third trigger signal comprising another second identifier (ID(CS(L_i_j))) of another second MPC representation (CS(L_i_j)) of another local machine learning model update (L_i_j) on the cluster of the aggregator (40); The fourth trigger signal includes another second identifier (ID(CS(L_i_j))).

7. The method (200) according to any one of claims 4 to 6, wherein the fourth trigger signal is sent (231) if a respective third trigger signal has been received by each of the plurality (C) of training clients (c_j) for a training iteration, the respective third trigger signal comprising a respective second identifier (ID (CS (L_i_j))) of a respective second MPC representation (CS (L_i_j)) of a respective local machine learning model update (Li_j) on a cluster of the aggregator (40); The fourth trigger signal includes each second identifier (ID(CS(L_i_j))).

8. The method (200) according to any one of claims 4 to 7, comprising: If a third identifier (ID (CS (G_(i+1)))) of a third MPC representation (CS (G_(i+1))) of an aggregated machine learning model (G_(i+1)) on the cluster of the aggregator (40) has been received, then checking (240) whether another training iteration for the aggregated machine learning model (G_(i+1)) should be performed; If another training iteration should not be performed, a third identifier (ID(CS(G_(i+1)))) is sent (241) to the owner (10) of the machine learning model.

9. The method (200) according to any one of claims 4 to 8, being executed within a Trusted Execution Environment (TEE).

10. The method (200) according to any one of claims 4 to 9, wherein the recording is carried out in a fully trusted manner on an external memory, optionally on a distributed ledger.

11. The method (200) according to any one of claims 4 to 10, wherein the orchestrator (20) causes (250) at least one training client (c_j) to download the aggregated machine learning model (G_(i+1)) and evaluate the quality of the aggregated machine learning model (G_(i+1)) based on the local test data of the training client; wherein at least one second test result is obtained, and the second test result is sent to the orchestrator (20); receiving (251) at least a second test result; evaluating (252) at least a second test result and a first test result obtained by a tester (50), wherein the tester is designed to evaluate the quality of the aggregated machine learning model (G_(i+1)) based on local test data of the tester (50), wherein the evaluation is based on MPC or homomorphic encryption, wherein an evaluation result is obtained; A predetermined action is performed (253) according to the evaluation result.

12. A computer-implemented method (300) for performing federated learning on a training client (c_j), comprising: downloading (320) a first MPC representation (CS(G_i)) of the machine learning model (G_i) from the cluster of the aggregator (40) based on the first identifier (ID(CS(G_i))) if a second trigger signal including the first identifier (ID(CS(G_i))) has been received and the identity and authority of the training client (c_j) has been verified on the cluster of the aggregator (40); converting (321) a first MPC representation (CS(G_i)) of the machine learning model into a local machine learning model according to a predetermined MPC protocol; Training (330) a local machine learning model based on local training data of the training client, wherein a local machine learning model update (L_i_j) is generated; converting (340) the local machine learning model update (L_i_j) into a second MPC representation (CS(L_i_j)) according to a predetermined MPC protocol; uploading (342) a second secret share of a second MPC representation (CS(L_i_j)) of the local machine learning model update (L_i_j) to the cluster of the aggregator (40), wherein a second identifier (ID(CS(L_i_j))) of the second MPC representation (CS(L_i_j)) of the local machine learning model update (L_i_j) on the cluster of the aggregator (40) is sent to the training client (c_j); A third trigger signal including a second identifier (ID(CS(L_i_j))) is sent (343) to the scheduler (20).

13. The method (300) of claim 12, being performed within a trusted execution environment (TEE).

14. A computer-implemented method (400) for performing federated learning for an aggregator (40), comprising: If the first secret share of the first MPC representation (CS(G_i)) is uploaded to the cluster of the aggregator (40), a first identifier (ID(CS(G_i))) of the first MPC representation (CS(G_i)) of the machine learning model (G_i) on the cluster of the aggregator (40) is sent (410) to the owner (10) of the machine learning model (G_i); If the second secret share of the second MPC representation (CS(L_i_j)) is uploaded to the cluster (40) of the aggregator, sending (420) a second identifier (ID(CS(L_i_j))) of the second MPC representation (CS(L_i_j)) of the local machine learning model update (L_i_j) on the cluster of the aggregator (40) to the training client (c_j)); If a fourth trigger signal is received, the local machine learning model update (L_i_j) is securely aggregated with at least one other local machine learning model update based on a predetermined MPC circuit (431), wherein a third secret sharing of a third MPC representation (CS(G_(i+1))) of the aggregated machine learning model (G_(i+1)) on the cluster of the aggregator (40) and a third identifier (ID(CS(G_(i+1)))) of the third MPC representation (CS(G_(i+1))) are obtained; A third identifier (ID(CS(G_(i+1)))) is sent (432) to the arranger (20).

15. The method (400) of claim 14, wherein the aggregated machine learning model (G_(i+1)) is based on weighted updates of local machine learning models (L_i_j).

16. The method (400) according to claim 14 or 15, comprising: sending (440) a third identifier (ID (CS (G_(i+1)))) of the third MPC representation (CS (G_(i+1))) to a tester (50), wherein the tester is designed to download the aggregated machine learning model (G_(i+1)) and evaluate the quality of the aggregated machine learning model (G_(i+1)) based on local test data of the tester (50), wherein the evaluation is based on MPC or homomorphic encryption, wherein a first test result is obtained, and the first test result is sent to the aggregator (40); Receiving (441) a first test result; The first test result is sent (442) to the scheduler (20).

17. A method (500) comprising: A computer-implemented method (100) for performing federated learning for an owner (10) of a machine learning model according to any one of claims 1 to 3; A computer-implemented method (200) for performing federated learning for an orchestrator (20) according to any one of claims 4 to 11; A computer-implemented method (300) for performing federated learning for a training client (c_j) according to any one of claims 12 to 13; and / or A computer-implemented method (400) for federated learning for an aggregator (40) according to any one of claims 14 to 16.

18. The method (100, 200, 300, 400, 500) according to any of the preceding claims, wherein the predetermined MPC protocol is based on fixed point numbers, floating point numbers and / or integers.

19. A device designed to carry out the method (100, 200, 300, 400, 500) according to any of the preceding claims.

20. A computer program, the computer program being designed to execute the method (100, 200, 300, 400, 500) according to any one of claims 1 to 18.

21. A data carrier or a signal comprising / encoding a computer program according to claim 20.