Reproduce “Your AI-Generated Image Detector Can Secretly Achieve SOTA Accuracy, If Calibrated” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Muli Yang, Gabriel James Goenawan, Henan Wang, Huaiyuan Qin, Chenghao Xu, Yanhua Yang, Fen Fang, Ying Sun and 2 more
Researchers from A*STAR and Xidian University demonstrate that existing AI-generated image detectors often exhibit substantial latent discriminative capabilities that are hindered by misaligned decision boundaries under distribution shift. A lightweight, post-hoc calibration framework is proposed to correct this bias, consistently improving detection accuracy by up to +16.16% across various benchmarks and maintaining robustness against image perturbations.
Key insights
- Many AIGI detectors possess robust latent discriminative features, but their effectiveness is undermined by systematic decision biases caused by train-test distribution shifts.
- Under specific, mild assumptions about distribution shift, the Bayes-optimal decision boundary for a classifier can be restored by applying a constant additive scalar correction to the model's logit outputs.
- Post-hoc calibration offers an efficient and universally applicable solution to enhance detector robustness and generalization without requiring expensive retraining or architectural modifications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.