SYSTEM AND METHOD FOR EVADING ADVERSARIAL ATTACKS ON DEEP NETWORK

摘要

Disclosed is a system and method for defending against hostile attacks on a deep network. The computer-implemented attack defense system includes at least one processor, the at least one processor including an attack defense module for defending against an attack against a deep network model using deep learning, and the attack defense module , In the training phase, a checkerboard is generated in the gradient of the input data and transferred to the deep network model, and the grid pattern generated in the input data is changed in the inference phase. It is transmitted to the deep network model.