SPOOFING-AWARE TRAFFIC SIGN RECOGNITION USING SELF-SUPERVISED ANOMALY DETECTION AND PREDICTIVE UNCERTAINTY

Traffic Sign Recognition (TSR) models are increasingly deployed in driver-assistance and autonomous driving pipelines, yet they remain vulnerable to spoofing and real-world manipulations such as stickers, overlays, graffiti, adversarial posters, and reflective film additions. These attacks can trigger confident misclassification—an especially hazardous failure mode because downstream planning and control subsystems often trust high-confidence predictions. This paper proposes a spoofing-aware TSR framework that integrates (i) self-supervised anomaly detection to identify distribution shifts and suspicious sign appearances without requiring labeled spoof examples, and (ii) predictive uncertainty modeling to quantify epistemic and aleatoric uncertainty and suppress overconfident errors. The approach combines representation learning from large-scale unlabeled driving data with reconstruction/consistency-based anomaly scoring and uncertainty-aware decision logic. We outline architecture components, training objectives, deployment strategy, and an evaluation protocol spanning physical tampering, synthetic spoofs, and natural corruptions. The framework aims to improve operational safety by enabling “detect–defer–downweight” behaviors: detect suspicious signs, defer to cautionary policies, and reduce the impact of unreliable predictions.