RollingEvidence: Autoregressive Video Evidence via Rolling Shutter Effect

Today, cameras are everywhere, generating vast video evidence essential for justice, security, and legal matters. However, advanced video manipulation and creation techniques threaten their authenticity, casting doubt on their reliability. Existing defensive measures, including data-driven models and digital watermarks, remain susceptible to adversarial and injection attacks. To address these challenges, we introduce RollingEvidence, an active system that utilizes the rolling shutter effect in CMOS cameras to embed real-time probes at the physical layer during recording, appearing as stripe patterns in video images, thus enhancing the integrity of probes and evidentiary content. RollingEvidence employs an autoregressive encoding scheme to produce compact, high-dimensional probes for later frames, incorporating previous frames and device-specific cryptographic keys. During verification, we design deep networks to decode probes from video and then locate tampered frames using exponential-min implication. Our theoretical analysis, along with a prototype and comprehensive experiments, demonstrate the efficiency of RollingEvidence in producing and verifying authentic videos.