R-PPDFL: A Robust and Privacy-Preserving Decentralized Federated Learning System

Federated Learning (FL), as an emerging distributed machine learning framework, which shares model gradients instead of raw data, can properly coordinate the contradiction between data sharing and data security under the guidance of laws and regulations and overcome the problem of “Data Silo”. However, the state-of-art federated learning schemes are still facing security challenges, such as single point of failure (SPOF), gradient privacy leakage, and byzantine attacks. To address the above issues, this paper proposes a robustness and privacy-preserving decentralized federated learning system (R-PPDFL). Specifically, a decentralized privacy-preserving federated learning framework based on the blockchain is designed and an improved multi-client functional encryption is proposed, which resolves the issues of SPOF and privacy leakage. Then based on functional encryption and cosine similarity we present a dense model detection method, which can properly defend the byzantine attacks in FL. Ultimately, it evaluates the proposed scheme by providing a theoretical analysis and conducting preliminary experiments on real datasets.