Multi-Scale Shapley Adaptation Pruning: Realizing Backdoor Defense in Brain-Computer Interface With Shapley-Value-Based Neural Network Pruning

Abstract

In the recent years, researchers made significant progress in electroencephalogram (EEG) classification tasks using deep neural networks, especially in brain-computer interface (BCI) systems. BCI systems rely on EEG signals for effective human-computer interaction, and deep neural networks have shown excellent performance in processing EEG signals. However, backdoor attack have a significant impact on the security of EEG-based BCI systems. In this paper, a novel multi-scale Shapley adaptation pruning (MSAP) method is proposed to solve the security problem caused by backdoor attack. In the proposed MSAP, the multi-scale Shapley segmented mapping method is used to accurately locate the backdoor weights. Subsequently, the cost function is utilized to adaptively prune the backdoor weights to ensure normal classification. Ultimately, the validity of the experiments is verified on the BCI competition public datasets (BCI-III-IVb, BCI-III-IVa, and BCI-IV-1a). The results show that the proposed MSAP method outperforms other pruning methods in defending EEG-based BCI systems against backdoor attack, maintaining a high baseline classification accuracy while reducing the attack success rate.