Strengthening Emotion Recognition Algorithms: A Defense Mechanism against FGSM White-Box Attacks

Emotion recognition algorithms have been widely used in various domains. However, the security of these algorithms has not received sufficient attention for a considerable time. In this paper, we propose a more attack-resistant emotion recognition model that can better withstand FGSM white-box attacks. The proposed model utilizes SE-ResNet for image feature extraction and integrates the extracted information into transformer modules and global depth convolutional layers. The anti-interference ability of the model is further enhanced through knowledge distillation, resulting in smoother convergence. Experimental results demonstrate that our proposed model achieves 74.95% accuracy on the FER2013 dataset and exhibits a certain level of resistance against FGSM interference.