Boosting adversarial example detection via local histogram equalization and spectral feature analysis

Deep neural networks (DNNs) are vulnerable to adversarial examples, which are created by adding imperceptible adversarial perturbations to images. To protect these DNNs, many powerful detection-based methods have been presented. However, they usually perform poorly on detecting adversarial examples with slight perturbations. In this work, we propose a novel dual-stream framework to enhance the detection of adversarial examples in deep neural networks. Our approach leverages local histogram equalization to amplify subtle adversarial perturbations, enabling clearer differentiation between adversarial and normal samples. We also extract and analyze both high-frequency and low-frequency information from images, exploiting the disruption of high-frequency features caused by adversarial perturbations. These features, alongside prediction confidence gradients based on low-frequency content, are fed into our detection network. Experimental results demonstrate that our method outperforms existing detection methods, particularly against attacks with minimal perturbations, achieving state-of-the-art performance on benchmark datasets. Code is available at https://github.com/LuyfDev/adversarial_detection.git.