Publicly Auditable Federated Learning With Privacy and Byzantine Robustness

Federated learning (FL) is a machine learning paradigm that can break the data silo effect, and it has been applied by lots of enterprises to train models that are deployed in inference services. However, existing FL schemes hardly support users who purchase inference services in auditing the training process, as the requirements for privacy and robustness restrict the realization of public auditability. This makes it difficult to allay users' doubts about whether the model is trained correctly. In this paper, we propose a publicly auditable and secure federated learning, which enables anyone to audit the correctness of model training securely. Specifically, we design a series of algorithms for generating the multiplication-related random vector, based on which we design a protocol that can generate the training record with masked privacy information for Byzantine-robust FL. The service provider can disclose the training record when the model is outputted, enabling users to reproduce the training process for auditing purpose without the need to know the privacy data of all parties involved in the FL. Meanwhile, to ensure model accuracy, we design an auditable incentive mechanism and a privacy-preserving federation member selection protocol to motivate workers to execute model training correctly. We analyze the security of our scheme and evaluate its efficiency. Compared to existing schemes requiring <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$O(R)$</tex-math></inline-formula> smart contract calls, our scheme requires only 3 calls. Also, experimental results under different settings show that the scheme is efficient and Byzantine-robust.