The vulnerability of neural networks to adversarial attacks has long been revealed. However, the structure of neural networks is not given due attention during the attack. The article deals with the impact of different parameters of a neural network on its resistance to adversarial attacks. The main purpose of this research is to determine which parameters increase resistance to attacks. The way by which neural networks can be compared has been proposed. Several neural networks were selected for comparison and a number of adversarial attacks were conducted on them. As a result, certain conditions were identified under which the attack took place over a longer time. It was also found that different changes in neural network parameters were required to protect against different attacks.
展开▼