Enhancing ATM Security with Deep Learning for Movement and Tampering Detection

Abstract

This paper proposes an advanced deep learning-based surveillance system aimed at enhancing security in Automated Teller Machine (ATM) environments by detecting human movement and physical tampering in real time. With increasing incidents of ATM-related crimes such as skimming, cash trapping, and forced break-ins, there is a critical need for intelligent systems capable of proactively identifying threats. The study introduces a modified YOLO (You Only Look Once) architecture, termed the YOLO Attenuation Model, which incorporates architectural optimizations to improve detection accuracy while reducing computational overhead, making it ideal for real-time applications in constrained ATM hardware settings. The model was trained and validated on a custom-curated dataset comprising various ATM-related scenarios, including normal usage, suspicious loitering, forced access attempts, and equipment tampering. Extensive experiments demonstrate that the proposed system achieves an outstanding mean Average Precision (mAP) of 0.99, significantly outperforming baseline object detection models in precision, recall, and inference speed. The high detection rate ensures rapid alert generation, enabling timely responses to security breaches. Furthermore, the system’s modular design allows for easy integration with existing ATM surveillance infrastructure. This research highlights the potential of deep learning to revolutionize ATM security by providing an automated, intelligent monitoring solution capable of reducing fraud and enhancing public safety.