Optimized Adversarial Example Generating Algorithm Based on Probabilistic Graph.

Deep learning technology is widely used in various fields. Once the deep model is attacked, it will cause huge economic losses and security problems. Therefore, the security of deep model has become a research hotspot. For attacking the depth model and detecting the robustness of the depth model, the adversarial examples is the core technology. And stAdv is one of the most advanced adversarial examples generation technologies. It has the advantages of a high success rate of attack and small visual distortion of adversarial examples, but it also has the problems of long generation time of adversarial examples and low efficiency. Aiming at the above-mentioned shortcomings, this paper improves the efficiency of adversarial examples generation algorithm, and proposes an optimized adversarial examples generation algorithm P&stAdv based on probabilistic graph. This method combines the steganography algorithm and CAM technique. P&stAdv perform attack evaluation on each pixel of the image, and introduces a method of obtaining the “appropriate point” of the image based on the steganography algorithm to generate the adversarial examples. Then the cost matrix of image modification is obtained and the image probabilistic graph is generated. Finally, use adversarial examples to attack images according to probabilistic graph. The experimental results show that the proposed algorithm has high efficiency. Compared with stAdv, the generation time of adversarial examples is reduced by 35%. This algorithm can modify the minimum number of pixels while ensuring a high success rate of attack. Moreover, it can make more effective attacks on complex texture images.