COUNTERATTACKS ON COMPUTER VISION SYSTEMS AND PROTECTION AGAINST THEM

Modern computer vision systems based on deep neural networks demonstrate the highest accuracy in object classification and recognition tasks. However, they turned out to be vulnerable to specially designed perturbations — adversarial attacks, which remain invisible to human perception, but can dramatically change the prediction of the model. This study is devoted to the systematic analysis of threats to the security of neural network models of computer vision. The paper presents a classification of attacks according to the level of available information about the model, discusses in detail the mechanisms for creating adversarial examples using gradient methods, and analyzes modern approaches to protection, including adversarial training and detection of abnormal input data. Special attention is paid to practical aspects: the results of testing the stability of popular architectures and quantitative indicators of the effectiveness of various protection methods are presented. The study confirms that the problem of adversarial attacks remains critically important for the deployment of reliable computer vision systems in real conditions.