Stealthy Adversarial Attacks: Decreasing Target Detection Performance through the Saturation Channel in HSV Color Space

With the widespread deployment of deep learning models in numerous security-critical applications, the security and robustness of these models have become a hot topic of research. Adversarial attacks, which inject imperceptible perturbations into input data, have revealed the vulnerabilities of these systems. This paper proposes an innovative adversarial attack strategy focusing on the saturation (S) channel in the HSV (Hue, Saturation, Value) color space. We hypothesize that by introducing carefully designed perturbations in the S channel, we can significantly reduce the performance of detectors without significantly altering the overall visual perception of the image. This study introduces subtle but effective adversarial perturbations in the S channel of the HSV color space, which are calculated based on the gradient of the detector in that channel. Through a series of experiments, we demonstrate that this method not only significantly reduces the accuracy of detectors but also has a high degree of stealth. Our research results indicate that the S channel is a potential weakness in target detection systems, and adversarial attacks can lead to a reduction in mean Average Precision (mAP) by up to 99% under specific conditions. Furthermore, we discuss the implications of these findings and propose possible defense measures against such attacks.