ENHANCING ADVERSARIAL ROBUSTNESS THROUGH STUDENT– TEACHER NETWORK COLLABORATION IN DEEP LEARNING MODELS

Abstract

Deep learning models are highly effective in complex decision-making tasks but remain vulnerable to adversarial attacks that exploit small perturbations in input data to mislead predictions. This study proposes a student–teacher network framework designed to improve adversarial robustness through knowledge distillation and adaptive defense learning. The teacher model, trained on clean and adversarially augmented datasets, transfers robust feature representations to the student network. The student learns to replicate the teacher’s resilient behavior while maintaining model generalization and accuracy. Experimental evaluations on benchmark image datasets, including CIFAR-10 and MNIST, demonstrate that the proposed method significantly enhances robustness against multiple adversarial attacks such as FGSM and PGD. The results indicate that the integration of student– teacher learning mechanisms can effectively mitigate vulnerability to adversarial perturbations while preserving model performance in standard conditions.