A STUDENT-TEACHER NETWORK APPROACH FOR DETECTING AND DEFENDING AGAINST ADVERSARIAL ATTACKS

Deep neural networks (DNNs) are highly vulnerable to adversarial attacks, where unnoticeable perturbations in input data can lead to severe misclassifications. Existing adversarial detection and defense techniques have a tendency to not be able to attain high classification accuracy for normal examples (NEs) but remain strong against adversarial examples (AEs). Here, we propose a new Student-Teacher Framework that effectively detects AEs and adopts a Restoration Attack (RA) mechanism to resist adversarial distortions without performance degradation on NEs. The teacher network is a normal classifier, and the student network forecasts teacher undistorted hidden layer features. The feature difference between the two networks serves as a detection score, which identifies adversarial perturbations. Upon detection, the restoration attack generates correction perturbations to restore adversarial feature space to the correct class regions and improve classification accuracy of AEs. Extensive experiments on CIFAR-10, CIFAR-100, and TinyImageNet datasets demonstrate that our framework achieves SOTA performance with high detection accuracy and significant robustness gains against white-box attacks, including PGD, CW, DeepFool, and AutoAttack. Our approach suppresses false positives and is resilient to adversarial errors without sacrificing classification accuracy on clean inputs. This robust and powerful framework provides an opportunity for secure deployment of DNNs in safety-critical systems.