Enhancing the Robustness of Deep Neural Networks using Deep Neural Rejection

Abstract

Adversarial examples are inputs that have been intentionally generated in order to deceive deep neural networks (DNNs) into generating inaccurate predictions. These instances endanger the security and safety of DNNs in real-world applications. To solve this problem, we present a new defense against adversarial instances based on Deep Neural Rejection (DNR). The DNR approach involves training a secondary model, referred to as the rejector model, to identify and reject inputs that are unlikely to produce correct predictions. The rejector model can be trained using adversarial examples as well as benign samples to learn the difference between the two types of inputs. If an input is rejected by the rejector model, it is assumed to be adversarial, and the primary model will not make a prediction on it. The experimental findings show that the DNR technique improves DNN resilience against hostile cases while retaining excellent accuracy on benign samples. Furthermore, by lowering the amount of samples that must be processed, the DNR technique can minimize the computational cost of DNNs.