Enhancement Of You Only Look Once Version 5 (Yolov5) Algorithm Applied in Nudity Content Detection for Video Chat Platforms

Abstract

The increasing prevalence of adversarial attacks poses significant challenges to object detection algorithms, including YOLOv5, which is widely used for its speed and accuracy in real-time applications. Adversarial manipulations, involving subtle and often imperceptible changes to input data, can result in severe misclassification or complete detection failure. This study investigates YOLOv5’s vulnerabilities to such attacks and proposes a Feature Squeezing Defense Method as a solution. The defense method incorporates Big Depth Reduction and Median Filtering techniques to effectively suppress adversarial perturbations while preserving the integrity of key object features. Extensive simulations revealed the substantial impact of adversarial attacks on YOLOv5’s performance, with detection confidence plummeting from 0.897 to 0.250 and Intersection over Union (IoU) reducing to 0.02, signifying critical failures in spatial alignment and classification accuracy. Following the implementation of the defense mechanisms, the model achieved significant recovery, with detection confidence improving to 0.904 and IoU alignment increasing to 0.94. Additionally, the proposed method enhanced the model's resilience, minimizing adversarial impact and ensuring reliable performance under hostile conditions. These findings highlight the importance of robust preprocessing techniques in addressing adversarial vulnerabilities and safeguarding object detection models. This research contributes to the development of secure and efficient systems for real-time applications, emphasizing their role in ensuring reliable and accurate performance across critical domains.