A new deepfake detector gets sharper precisely when the fakes get blurrier.
Researchers built a system called BAM-Net that converts an image into a frequency-domain map (a discrete cosine transform, the same math behind JPEG compression) and learns which frequency bands carry the telltale traces of manipulation. It dynamically boosts those bands and dampens the noise, then merges that signal with a standard spatial-analysis network to make a final call. Tested on three standard benchmarks - FaceForensics++, Celeb-DF, and DFDC - the model posted state-of-the-art results, including on forgery types and compression levels it never saw during training.
That cross-compression detail is the real news. Most deepfake detectors are trained on clean video and fall apart once content gets re-encoded for upload to social platforms, which is exactly what happens to nearly all fake video circulating online. A method built to counteract compression loss rather than assume pristine footage is targeting the actual failure mode, not the lab condition.
Still, this is a benchmark paper, not a deployed tool, and benchmark generalization has a habit of looking better on paper than against next year's generation of forgeries.