Generate Adversarial Examples Combined with Image Entropy Distribution

With the rapid development of artificial intelligence technology, artificial intelligence-based systems and applications have shown explosive growth, bringing great convenience to people's lives. As a representative of artificial intelligence, deep learning, whose excellent performance in the classification task has been proved, can obtain high-precision classification models by effectively training from large amounts of data. However, recent studies have shown that deep neural networks are vulnerable to attacks. By carefully constructing input data, Deep Neural Networks can export the output results expected by the attacker. In general, it is difficult to detect the difference between the carefully constructed data and the original data, but it can induce the neural network to output wrong results. This kind of attack is called adversarial example attack, and the carefully constructed data used to deceive deep learning is called adversarial examples. This paper proposes a generation method of adversarial example that combines the distribution of image entropy. The areas with high entropy values in the image tend to have complex tones, so the added perturbation cannot be noticed by naked eyes easily, and experiments have proved that the adversarial examples generated by adding perturbations in this area are more aggressive. This paper proposes that the greater the entropy in a region of an image, the greater the weight assigned to the position of the Adversarial Perturbation; the smaller the entropy in a region of the image, the lower the weight is assigned to the position of the Adversarial Perturbation, so that an adversarial example with lower disturbance and less noticeable is generated, and at the same time, the adversarial example is more aggressive.