New AI Framework Uses Adaptive Frequency Analysis to Catch Deepfakes Across Different Compression Standards

Face forgery detection—the challenge of identifying digitally manipulated or deepfake videos—remains difficult because detection systems often fail when encountering new manipulation techniques or compressed media. A new approach tackles this generalization problem by teaching machines to focus on the right frequency patterns in video frames.

The Core Challenge#

Modern deepfake detectors face two major obstacles. First, they struggle to adapt when faced with manipulation methods they haven't encountered during training. Second, real-world video compression degrades the subtle artifacts that typically signal forgery, making detection harder. Existing systems rely on rigid frequency filters or broad band divisions that can't learn which spectral features matter most for the task at hand.

How BAM-Net Works#

Researchers at arXiv have proposed the Band-Attention Modulation Network (BAM-Net), which takes a learnable approach to frequency analysis. The system converts video frames into the Discrete Cosine Transform (DCT) domain—a standard representation used in image compression—and then dynamically reweights different frequency bands to highlight forgery-related patterns while suppressing noise.

The key innovation is a Band-Attention Modulation (BAM) mechanism that:

• Examines frequency components organized along anti-diagonals in the DCT spectrogram
• Learns which bands carry the most diagnostic information for detecting manipulation
• Reweights these bands adaptively, simulating a kind of "inverse compression" that recovers lost detail

This modulated frequency information is then combined with spatial domain analysis using a lightweight neural backbone equipped with distance-decayed attention, enabling comprehensive feature extraction across both domains.

Experimental Results#

Testing on three major datasets—FaceForensics++, Celeb-DF, and DFDC—showed that BAM-Net achieves top-tier performance. More importantly, the system generalizes well across challenging scenarios:

• Cross-dataset transfer: Detection accuracy holds up when models trained on one dataset are applied to another
• Compression robustness: The approach remains effective even as videos undergo different compression levels
• Cross-manipulation generalization: The system can identify forgeries created with manipulation techniques it hasn't seen before