Researchers Propose Agentic Framework Using Electric Network Frequency to Authenticate Video Against Generative AI Models

Generative AI models produce highly convincing video content, making conventional artefact-based learned detectors unreliable when facing unseen generation techniques. To address this limitation, researchers have developed an agentic forensic framework centered on Electric Network Frequency (ENF) signals. These signals represent ambient mains-driven power grid fluctuations naturally captured as audio hums or video flicker during recording, providing a physical layer of data independent of generative model control. The proposed system employs a unified two-channel detector that extracts ENF signatures from audio and rolling-shutter video using a shared harmonic estimator network. To solve the shortage of labeled training data, the authors introduced a synthetic-injection training method powered by an Ornstein-Uhlenbeck drift model. Crucially, the research revealed that generated videos can naturally exhibit ENF-band flicker, meaning presence alone does not prove authenticity. Instead, verification relies on matching the signal against an independent reference recording from the power grid. To balance computational overhead and accuracy, the framework utilizes a two-axis agentic orchestrator that fuses recording authenticity with content manipulation evidence. Operating under an asymmetric rule, this orchestrator escalates execution to resource-intensive detection pipelines only when necessary. On held-out real recordings, the framework demonstrated high accuracy while matching the performance of a flat model ensemble at reduced operational cost.