AV-FABLE: Detecting Audio-Visual Manipulation in Long-Form Videos
benchmarks
| Source: HF Papers | Original article
Researchers introduce AV-FABLE, a forensic tool that detects localized audio‑visual manipulations in long‑form videos, addressing gaps in current deepfake benchmarks that focus on short clips.
A new benchmark called **AV‑FABLE** (Audio‑Visual Manipulation Forensics in Long‑Form Videos) has been released to address a growing blind spot in deep‑fake detection. Researchers note that many AI‑generated alterations in lengthy videos are confined to brief, isolated events, leaving the surrounding footage untouched. Traditional forensic tools that classify an entire video as real or fake therefore miss these subtle tamperings.
AV‑FABLE departs from earlier datasets that concentrate on short, talking‑head clips. It presents a “dual‑timeline” suite of open‑scene, long‑form videos in which manipulations are sparse and can affect either the visual stream, the audio stream, or both. The benchmark also introduces a frozen‑feature readout model, NFSR, which outperforms zero‑shot audio‑visual large language models in pinpointing the exact moments of tampering. NFSR maintains its edge across three out‑of‑distribution scenarios, including transfer to a separate long‑video forensic benchmark that was held out during training.
The release is significant because it equips researchers and forensic analysts with a realistic testbed for detecting localized deepfakes that could be weaponised in political propaganda, disinformation campaigns, or illicit content manipulation. By exposing the limits of video‑level classification, AV‑FABLE pushes the field toward finer‑grained, multimodal analysis that mirrors how real‑world forgeries are likely to be crafted.
The authors will make the benchmark, code and model weights publicly available, inviting the community to build on the baseline. Watch for early adopters integrating AV‑FABLE into academic competitions and commercial verification pipelines, and for follow‑up studies that expand the dataset to cover more diverse domains such as live‑streamed events and user‑generated content.
Sources
Back to AIPULSEN