Core Viewpoint - Ant Group's AIoT technology team has developed an innovative active video verification system called RollingEvidence, which utilizes the rolling shutter effect of cameras to embed high-dimensional physical watermarks in videos, effectively countering deepfake and video tampering attacks [2][4][6]. Group 1: Innovation and Technology - RollingEvidence transforms the "defect" of CMOS cameras into a security advantage by injecting rolling stripe detection signals into each video frame, creating a "digital pulse" for real-time verification [4][6]. - The system employs a self-regressive encryption mechanism to ensure that the content is non-falsifiable and tampering is traceable, enhancing the accuracy and security of video verification compared to traditional passive recognition technologies [4][6]. - The system's architecture includes a specialized deep neural network that extracts stripe features and decodes probe information, allowing for precise identification of tampered frames [21][28]. Group 2: Performance and Application - RollingEvidence has been validated through theoretical analysis, prototype implementation, and extensive experiments, demonstrating its effectiveness in generating and verifying trustworthy video evidence [6][46]. - The system is applicable in critical scenarios such as notarization, identity verification, and judicial evidence collection, addressing the challenges posed by advanced AI video generation technologies [6][46]. - Experimental results indicate that RollingEvidence can accurately detect most tampering behaviors without misjudging normal videos, achieving high accuracy rates across various testing scenarios [38][40][41]. Group 3: Experimental Results - The system's tampering detection performance was evaluated through two sets of experiments, showing it can accurately identify frame insertion, deletion, and modification, as well as face swapping and lip-sync detection [37][38]. - In various scenes, the system achieved an accuracy rate of up to 99.84% with a false rejection rate (FRR) of 0.00% and a false acceptance rate (FAR) as low as 0.22% [38]. - The performance of the verification submodule was also assessed, demonstrating high precision in stripe extraction and excellent denoising effects, even under varying background and lighting conditions [44].