Problem

Deep image forgery detection faces challenges in generalizability and robustness due to frequency bias in DNN detectors.

Method

The authors propose a two-step frequency alignment method involving Spectral Magnitude Rescaling and Reconstructive Dual-domain Calibration to remove frequency discrepancies between real and fake images.

Results

The frequency alignment method effectively improves detector generalization and robustness and can also be used as a strong black-box attack.

Implications

This study enhances the reliability of deep forgery detectors and opens avenues for future research in other image domains and modalities.