Enhanced slider CAPTCHA recognition for vulnerability assessment / Wan Xing, Juliana Johari and Fazlina Ahmat Ruslan

This paper addresses the limitations of existing slider CAPTCHA recognition methods by proposing an enhanced approach based on the YOLOv5 model and introduces a novel evaluation metric, mean Relative Offset (mRO), which more accurately assesses the x-coordinate of gap location prediction compared to traditional mean Average Precision (mAP). Furthermore, a novel Offset-based Intersection over Union (OIoU) loss function is proposed, specifically designed to prioritize the accuracy of horizontal displacement, crucial for slider CAPTCHA detection. The study also presents Fixed Quantity Prediction-based Non-Maximum Suppression (FQP-NMS), a modified NMS algorithm ensuring a fixed number of predicted bounding boxes, addressing the variability inherent in standard NMS. Experiments demonstrate that the proposed OIoU and FQP-NMS, when integrated with YOLOv5, significantly improve both mAP and mRO, particularly on challenging SliderCAPTCHA datasets. The incorporation of lightweight attention mechanisms, such as ECA, further enhances the model's performance and robustness, especially when combined with VGG19 and other efficient architectures. The results indicate that the proposed system provides a more accurate and robust solution for slider CAPTCHA recognition.