Research on adversarial robustness properties of image classification networks based on deep vision

In response to the problem of significant performance decline of existing deep learning-based intelligent recognition algorithms under adversarial sample attack conditions, this research investigates the intrinsic mechanisms and description methods of adversarial samples. Quantitative linear characteristic analysis is conducted on sub-operations of convolutional neural networks, a model is established to compute the incremental output corresponding to perturbed inputs of suboperations, and the internal mechanism of adversarial sample generation is explored. Using the fast gradient descent method, sensitivity coefficients and offset coefficients are introduced in RestNet networks to establish a relationship model between input perturbations and outputs. The linear characteristics in high-dimensional space are demonstrated to be the cause of adversarial sample generation. Finally, using the projection gradient descent method, a relationship model is established between the number of iterations and outputs to solve the mapping relationship between sensitivity coefficients and the number of iteration attacks. This provides guidance for the design of deep learning attack-defense algorithms.