8
Just got my first deepfake detection script to work after a solid week of banging my head against it
I wanted to build a simple tool to flag potential deepfakes in local news comment sections, thinking it would be a weekend project. The core idea was to check for weird pixel flow and audio sync mismatches. I hit a wall right away because the open source models I found were trained on high quality fakes, not the grainy, compressed junk people actually post. I spent three full days just trying to tweak the sensitivity thresholds so it wouldn't flag every badly encoded video as fake. My final fix was stupid simple: I added a basic check for the video's original bitrate and source before even running the deepfake scan. It took seven days total when I thought it would take two. Has anyone else tried to make a detection tool for low quality stuff and found a better way to filter the noise first?
3 comments
Log in to join the discussion
Log In3 Comments
victor7798d ago
How many false positives are you still getting even after the bitrate check? I tried something similar last year and found compression artifacts from screen recordings would still trip it up every time.
2
oliverhernandez8d ago
Honestly, how do you even tell the difference sometimes? I mean, I was looking at some old family videos my dad digitized, and the quality is so bad it looks like it was recorded through a screen door. It's all blocky and weird, but it's the original tape, you know? Makes you wonder what counts as an artifact versus just a really bad source.
5
taylorshah8d ago
Man, that bitrate check is smart. I ran into the exact same wall with compression junk. What finally worked for me was adding a simple face detection step first. If the tool can't find a clear face to check, it skips the deepfake scan completely. It cut my false positives way down because a lot of the noisy garbage videos don't even have a proper face in frame.
1