Abstract:

Most of the existing physical adversarial attack methods are limited to planar patches, and even the adversarial samples that can perform multi-angle attacks suffer from insufficient robustness, insufficient generalization, and a large gap between the attack effects in the digital and physical domains. A white-box vehicle adversarial texture generation method is proposed based on this: add images with different brightness and contrast in the training dataset, and add noise that simulates the real environment on the texture generated after each training epoch, use the Bayesian optimization algorithm to compute the weights of the different loss terms, and finally add a regularization term to reduce the overfitting of the model. In response to the problem that the model and the target of the existing dataset cannot be completely overlapped, an inpainting method is proposed for repairing images to reduce the gap between the digital simulation and the real shot. Digital simulation experiments and physical world experiments show that the proposed algorithm achieves a higher attack success rate and lower precision rate compared to existing adversarial texture generation algorithms.

Key words: adversarial attack, physical attack, texture generation, white-box attack