Robust Federated Learning Algorithm Based on Chaotic Encryption and Dual-Server Detection

Federated learning allows users to collaboratively train models, but studies have shown that local model parameters may leak users’ privacy. Furthermore, the global model in federated learning could be compromised by Byzantine attacks. To address these issues, a robust federated learning algorithm based on chaotic encryption and dual-server detection (CDRFL) is proposed. First, CDRFL introduces a lightweight chaotic encryption method, where ciphertexts have a unique property of mutual cancellation, allowing the plaintext global model parameters to be derived without decryption. Moreover, CDRFL allows users to encrypt with different keys, further reducing the risk of privacy leakage. Next, a dual-server detection scheme is designed to defend against Byzantine attacks while preserving user privacy. Lastly, a reputation score mechanism with a forgetting function is proposed to fairly distinguish Byzantine participants. Experimental results demonstrate that, compared with popular robust aggregation schemes, CDRFL improves testing accuracy by 3.26%–9.06%. When compared to existing chaotic encryption federated learning schemes, CDRFL enhances the ability to defend against Byzantine attacks and reduces time and communication overhead by 73.63% and 95.40%, respectively.