Working on a hackathon project that flags potentially manipulated short-form videos. the desktop deepfake detectors ive tried fall apart on 15-60 second clips with low resolution. is anything actually useful at this format
sensity and Reality Defender both publish that they handle short form, but in my (limited) tests sensity was more accurate on tiktok-style clips. they both have apis for hackathon use. the issue with short form isnt the model, its that compression destroys the artifacts the detectors look for. expect 70% accuracy ceiling at best, no matter the tool.
On sensity. also keep in mind facial-area detectors do better on clips where a face is dominant on screen. anything wider gets harder.
@RustyCircuitX 70% ceiling is good to know, we were targeting 90% which would have been a fool’s errand.
If you have any agency to design the project, narrow the scope to ‘face-dominant’ clips for v1. the platform-level ceiling is real.
academic literature on short-form deepfake detection is sparse precisely because the compression destroys what you’d measure. the closest applied work is in TikTok-specific moderation tooling, which is proprietary. expect a 70% ceiling for the foreseeable future.
If you can capture pre-compression source (eg direct from creator instead of post-platform) the accuracy ceiling goes up significantly. but thats only available to platforms themselves, not third-party tools.