PSFL: Ensuring Data Privacy and Model Security for Federated Learning

The integration of blockchain-based federated learning (BFL) and Industry 4.0 utilizes intermediate models to execute task deployment and result acceptance, effectively solving the problems of data barriers and data resource waste in Industry 4.0. However, the BFL ecosystem is susceptible to poisoning and inference attacks that undermine data privacy and model security. In this article, we propose PSFL, a robust federated learning framework that guarantees both data privacy and model security. Specifically, we design a cross-validation algorithm where numerous participants conduct a thorough assessment of the user's contribution growth rate in the current round. This approach proves effective in identifying Byzantine attackers engaged in malicious activities within the system. Furthermore, we propose a lightweight multireceiver signcryption mechanism employing secure key distribution, which significantly minimizes resource overhead. Finally, the security of PSFL is proved based on the random oracle model. Empirical assessment affirms the effectiveness and practicality of PSFL, even with different proportions of malicious users, PSFL's performance is 10%-20% higher than trimmed mean and M-Krum. In summary, PSFL improves the model accuracy and the security of the model transmission process in scenarios involving edge node poisoning, which demonstrates that PSFL can be well adapted to Industry 4.0.