Motion-sensor-based anti-spoofing in a mobile identity verification application

Abstract

Remote identity verification solutions face new challenges: with advances in generative AI, a single facial image of a victim can suffice for adversaries to synthesize lifelike video models and conduct deepfake video-injection attacks. A practical mitigation to such attacks would be to include additional data modalities in the verification process, reducing reliance on the video feed alone, complicating potential attacks. On consumer mobile devices, one widely available modality would be motion sensors, for example, accelerometers and gyroscopes.

This thesis examines the potential of combining motion-based sensor data collected during Candour ID's selfie-capture phase with state-of-the-art time-series classification and anomaly-detection methods. For this purpose, a data set was collected from 30 subjects (10–15 samples each), totaling 375 time-series sequences of multi-sensor data sampled during selfie captures. Three distinct, comprehensive benchmarks were conducted: (i) multi-class classification to identify individual users, (ii) an anomaly-detection evaluation to flag possible spoofing attacks or unusual behavior, and (iii) a 10-shot, semi-supervised one-class learning -based verification (anomaly detection) to verify individual users without relying on a multi-class classifier. The results indicate that a selfie-capture-centered motion data analysis check shows promise as a complementary security layer supporting Candour's face-matching engine.