Parameter-Efficient Adversarial Example Detection and Robustness Enhancement Utilizing Optimized Reverse-Cross Entropy

Reverse Cross-Entropy (RCE) has shown promise in enhancing adversarial example detection (AED) and robustness capabilities of deep neural network (DNN) models under adversarial attacks. However, high retraining costs and unstable convergence on large-class datasets have limited RCE loss function's adoption and application. In this work, we refine RCE with temperature scaling and weighted focusing (RCE-TF) to stabilize training on larger tasks. Additionally, we investigate the use of bottleneck residual adapter blocks to efficiently enhance the robustness of pretrained models. Experiments on CIFAR-10 and CIFAR-100 demonstrate that our proposal scales RCE to large dataset classification problems with higher average accuracy compared to conventional loss function. We also achieve superior AED performance and improved robustness up to 10% under multiple adversarial attacks and with specific adapter settings, without prior knowledge to these attacks.