An incremental adversarial training method enables timeliness and rapid new knowledge acquisition

Adversarial training is an effective defense method for deep models against adversarial attacks. However, current adversarial training methods require retraining the entire neural network, which consumes a significant amount of computational resources, thereby affecting the timeliness of deep models and further hindering the rapid learning process of new knowledge. In response to the above problems, this article proposes an incremental adversarial training method (IncAT) and applies it to the field of brain computer interfaces (BCI). Within this method, we first propose a deep model called Neural Hybrid Assembly Network (NHANet) and then train it. Then, based on the original samples and the trained deep model, calculate the Fisher information matrix to evaluate the importance of deep neural network parameters on the original samples. Finally, when calculating the loss of adversarial samples and real labels, an Elastic Weight Consolidation (EWC) loss is added to limit the variation of important weights and bias parameters in the Neural Hybrid Assembly Network (NHANet). The above incremental adversarial training method was applied to the publicly available epilepsy brain computer interface dataset at the University of Bonn. The experimental results showed that when facing three different attack algorithms, including fast gradient sign method (FGSM), projected gradient descent (PGD) and basic iterative method (BIM), the method proposed in this paper achieved robust accuracies of 95.33%, 94.67%, and 93.60%, respectively, without affecting the accuracy of clean samples, which is 5.06%, 4.67%, and 2.67% higher than traditional training methods respectively, thus fully verifying the generalization and effectiveness of the method.