Abstract
•We introduce a secure aggregation scheme that processes each client’s encrypted local model in parallel, safeguarding parameter privacy while tolerating dropouts.•To prevent malicious client sends wrong secret shares, we adopt verifiable secret sharing (VSS) to validate every share and guarantee correct global aggregation.•Vector homomorphic hashing (VHH) is further employed to stop intermediate entities from tampering with the server and to certify the integrity of the returned aggregate.•Security analysis confirms resistance to all considered attacks, and experiments on MNIST and CIFAR-10 demonstrate competitive performance.
[Display omitted]
The existing secure aggregation schemes for federated learning (FL) have made some progress in protecting the privacy of client local model parameters and verifying the correctness of aggregation results. They usually use secret sharing (SS) technology to deal with dropout clients, but there are two limitations: one is that the computational cost of handling the dropouts and verifying the aggregation results is large. Second, it does not consider the problem of false shares, which affects the robustness of the global model. Therefore, this paper proposes a secure aggregation scheme with verifiability and robustness for privacy-preserving federated learning (PPFL) to solve the above problems. Firstly, the local model parameters are encrypted by mask technology to ensure the confidentiality of the client’s local model parameters. Secondly, by adopting Verifiable Secret Sharing (VSS) technology, it can not only efficiently handle dropout clients but also actively filter out forged shares submitted by malicious clients, thereby ensuring the robustness of the aggregation process. Finally, we introduce the Vector Homomorphic Hashing (VHH) mechanism to replace the original complex verification algorithm, thereby achieving accurate verification of the aggregation results in a lightweight manner. Security analysis shows that the scheme can protect client privacy and training integrity. Experiments show that compared with the existing scheme without considering fault sharing detection, the computational efficiency of the proposed scheme is improved by about 34%. At the same time, when there are 30% malicious clients, the computational overhead also tends to be stable, which is more suitable for resource-constrained scenarios such as mobile devices and cross-organizational collaboration.