Evading generated-image detectors: A deep dithering approach

•A deep dithering model is proposed to evade general generated-image detectors in a black-box and end-to-end way.•Effective VQ-VAE based generated-image simulator is employed for self-supervised anti-forensic training data preparation.•A new contrastive loss in fraction form is adopted for deep anti-forensic model training.•The proposed method achieves superior anti-forensic performances on various generated images in terms of visual fidelity and undetectability.