Frequency bias matters: an in-depth investigation into robust and generalizable deepfake image detection

For image forgeries driven by AI-generated models such as GANs, researchers analyze deepfake image detection from a frequency perspective and find that the frequency bias of DNN forgery detectors may be a cause of their generalization and robustness problems. Based on this, the paper proposes a two-step frequency alignment method to eliminate the frequency difference between real images and forged images. This method has dual use: it can serve as a strong black-box attack against forgery detectors in anti-forensics scenarios, and as a general defense in forensics scenarios to improve detector reliability. The authors developed corresponding attack and defense implementations and demonstrated effectiveness in multiple experimental settings involving 12 detectors, 8 forgery models and 5 metrics.