Resistance researching of YOLO object detectors to adversarial attacks

In physical world, object detectors encounter various interferences and noises, as well as malicious modifications that can significantly affect their performance. Such changes can lead to the removal of compromised detectors from their normal operating modes, which actualizes the tasks of monitoring existing and newly created YOLO models for resistance to adversarial attacks. Purpose: The aim of the study is to assess the resilience of intelligent object detectors to complex adversarial attacks during transfer training. Methods: Machine learning and adversarial defense methods were used. Results: the results of evaluating the performance of various YOLO family models against adversarial attacks aimed at gradients have shown that attacks on gradients of the confidence loss function are the most effective, leading to degradation of model performance from 20% with a small attack budget to 100% with large budgets of deep attacks on gradients. Practical relevance: the tactics of a comprehensive attack on model gradients proposed in the paper substantiate the expediency of using competitive learning mechanisms as a mandatory protection measure, as well as additional measures such as training and operating verified YOLO models in trusted execution environments. Discussion: It is not possible to counteract deep adversarial attacks on gradients with any one method of defence, the necessary and sufficient set of tools is determined by the specifics of the model operation.