Generating Adversarial Examples for Robust Deception against Image Transfer and Reloading

Adversarial examples play an irreplaceable role in evaluating DNNs' security and robustness. It's essential to understanding the adversarial examples' effectiveness to utilize them for model improvement. In this paper, we explore the impact of input transformation on adversarial examples. First, we discover a new phenomenon. The process of RELOAD or TRANSFER may deactivate adversarial examples' malicious functionality. The reason is that these processes would reduce pixel precision, which counters the perturbation added by the adversary. We validate this finding on different mainstream adversarial algorithms. Second, we propose a novel Confidence Iteration method, which can generate more robust adversarial examples. The key idea is to set the confidence threshold and add the pixel loss caused by image reloading or transferring into the calculation. We integrate our solution with existing adversarial algorithms. Experiments indicate that such integration can significantly increase the adversarial attacks' success rate.