RADNN: ROBUST TO IMPERCEPTIBLE ADVERSARIAL ATTACKS DEEP NEURAL NETWORK

This paper presents the RADNN algorithm. The RADNN is a robust to imperceptible adversarial attack algorithm that uses the concept of data density and similarities to detect attacks on real-time. Differently from traditional deep learnings that need be trained on the attacks to be able to detect, RADNN has a mechanism that detects data patterns changes. In order to evaluate the proposed method, we considered the PerC attacks and a 1000 images from the Imagenet dataset. The RADNN could correctly identify 97.2% of the attacks.