A backdoor attack defense scheme in federated learning

Federated learning allows for the effective flow of data value without leaving the local environment. It is considered an effective method for balancing data sharing and privacy protection in the context of the Internet of Things. However, federated learning systems are susceptible to Byzantine attacks and inference attacks, resulting in compromised system robustness and data privacy. A secure federated learning algorithm DP-FLABA based on gradient local density is proposed to address the Byzantine attack problem in the federated learning framework. This algorithm calculates the local density of each client gradient to determine whether the client is a Byzantine client, and then designs corresponding weights for weighted aggregation based on the local density of the secure gradient. At the same time, in response to privacy leakage issues, adaptive pruning and denoising techniques are adopted to protect model parameters. We have ensured data privacy while considering model robustness and accuracy, and theoretically analyzed the privacy of the protocol. The simulation results show that this scheme can effectively resist Byzantine attacks and inference attacks.