Adversarial Robustness Evaluation with Separation Index

The paper introduces a method to assess the robustness of deep neural networks against adversarial attacks. It employs a geometric-based separation metric called the Separation Index, which measures the distance between data points with distinct labels within the latent space of variational autoencoders utilized for classification tasks. The Separation Index quantifies the degree of data separation by comparing each data point with its neighboring data points. A higher value signifies greater separation between different classes, thus ensuring enhanced robustness. This approach yields dependable results when confronted with gradient-based adversarial attacks, including FGSM, R-FGSM, MI-FGSM, and PGD, under both white-box and black-box conditions.