IPCADP-Equalizer: An Improved Multibalance Privacy Preservation Scheme against Backdoor Attacks in Federated Learning

Abstract

Although there are some protection mechanisms in federated learning, its training process is still vulnerable to some powerful attacks, such as invisible backdoor attacks. Existing research work focuses more on how to prevent attacks in distributed training scenarios and improve the security of the FL training process, but it lacks consideration of utility and robustness, especially when the learning model of FL suffers from stealth backdoor attacks. This paper proposes an improved FL defense scheme IPCADP based on user-level differential privacy and variational autoencoders technology. The scheme can control and protect the privacy attribute of the image and can also eliminate the triggers that exist in the poisoned image. The experimental results show that compared with some existing defense schemes, IPCADP can defend against invisible backdoor attacks and improve the classification accuracy of the main task, while mitigating the impact of attacks on model robustness and stability. To a certain extent, the balance and unity of security, utility, and robustness are realized.