Real-Time Micro-Expression and Eye-Gaze Analysis forBehavior-Based ATM Fraud Detection

Abstract

Automated teller machines (ATMs) remain one of the most frequently exploited nodes in the retail banking ecosystem, with fraud typologies ranging from card skimming and shoulder-surfing to coerced withdrawals under physical duress. Conventional fraud-prevention layers — PIN verification, card-chip validation, and transaction-pattern analytics — operate on transactional metadata and are structurally blind to the physical and psychological state of the person standing at the terminal. This paper proposes a real-time, camera-based behavioral analysis framework that fuses facial micro-expression recognition with eye-gaze trajectory analysis to detect anomalous affective and attentional states that correlate with fraudulent or coerced ATM activity. The system, termed MER-Gaze Net, combines a lightweight 3D convolutional network for spatio-temporal micro-expression encoding with a recurrent gaze-trajectory encoder, fused through a cross-modal attention layer that outputs a continuous risk score alongside a discrete behavioral class. Evaluated on a composite dataset of 9,200 annotated ATM interaction sessions spanning genuine, coerced, impersonation, and skimming-staging behaviors, the proposed model achieves 95.7% classification accuracy and an AUC of 0.978, outperforming unimodal and late-fusion baselines by 5.4 to 24.3 percentage points. We further characterize the accuracy-latency trade-off at deployable frame rates and the system's robustness under degraded illumination, showing that infrared-assisted low-light operation recovers most of the performance lost under sub-50-lux conditions. The paper closes with a discussion of deployment architecture, privacy-preserving edge-inference design, and the ethical constraints that must govern any system that infers psychological state from involuntary facial signals.