A Novel Convolutional Autoencoder Framework to Mitigate FGSM-Based Adversarial Threats in Face Recognition

Abstract

Face recognition systems have achieved remarkable performance in numerous real-world applications, but they remain vulnerable to adversarial attacks that subtly manipulate input data to deceive models. One of the most prevalent forms of attack is the Fast Gradient Sign Method (FGSM), which generates adversarial images that can significantly reduce recognition accuracy. In response to this challenge, we propose a novel convolutional autoencoder framework specifically designed to defend face recognition models from FGSM-based white-box adversarial threats. The proposed method leverages the denoising capabilities of convolutional autoencoders to reconstruct clean facial features from perturbed inputs, thereby restoring recognition performance. Through extensive experimentation on benchmark face datasets such as LFW and YaleB, we evaluate the robustness and effectiveness of our defense mechanism. The results demonstrate a substantial improvement in classification accuracy and resilience under attack conditions, indicating the potential of autoencoders as an efficient and scalable defense strategy.