PID-Based Approach to Adversarial Attacks

Adversarial attack can misguide the deep neural networks (DNNs) with adding small-magnitude perturbations to normal examples, which is mainly determined by gradient of loss function respect inputs. Previously, various strategies have been proposed enhance performance adversarial attacks. However, all these methods only utilize gradients in present and past generate examples. Until now, trend change future (i.e., derivative gradient) has not considered yet. Inspired classic proportional-integral-derivative (PID) controller field automatic control, we propose a new PID-based approach for generating The past, are our method, correspond components P, I D PID controller, respectively. Extensive experiments consistently demonstrate that method achieve higher success rates exhibit better transferability compared state-of-the-art gradient-based Furthermore, possesses good extensibility be applied almost available