Leveraging Model Distillation as a Defense Against Adversarial Attacks Based on Deep Learning

Adversarial attacks on deep learning models threaten machine learning system security and reliability. The above attacks use modest data alterations to produce erroneous model results while being undetected by humans. This work suggests model distillation to prevent adversarial perturbations. The student model is taught to emulate the teacher model in model distillation. This is done using teacher model soft outputs. Our idea is that this strategy organically strengthens the student model against adversarial assaults by keeping the teacher model's essential knowledge and generalization capabilities while reducing weaknesses. Distilled models are more resilient to adversarial assaults than non-distilled models, according to experiments. These models also perform similarly on undamaged, uncorrupted data. The results show that model distillation may be a powerful defense against machine learning adversaries. This method protects model resilience and performance.