DNN Attack detection using contrastive explanations

Explainability Driven Online Attack Detection in Deep Neural Networks: First, a DNN model is trained. Next, the trained model is used to generate functional patterns. Thereafter, the model is injected with faults for attack simulation. Finally, the pre-computed patterns are applied as input to the perturbed model for detecting attacks.