Your AI-Generated Image Detector Can Secretly Achieve SOTA Accuracy, If Calibrated

Despite being trained on balanced datasets, existing deepfake detectors often exhibit systematic bias at test time, frequently misclassifying fake images as real. We hypothesize that this behavior stems from distributional shifts in fake samples and implicit priors learned during training. Specifically, models tend to overfit to superficial artifacts that do not generalize well across different generation methods, leading to a misaligned decision threshold when faced with test-time distribution shifts. To address this, we propose a theoretically grounded post-hoc calibration framework based on Bayesian decision theory. Specifically, we introduce a learnable scalar correction to the model’s logits, optimized on a small validation set from the target distribution while keeping the backbone frozen. This parametric adjustment compensates for distributional shifts in model output, realigning the decision boundary without requiring ground-truth labels. Experiments on challenging benchmarks show that our approach significantly improves robustness without retraining, offering a lightweight and principled solution to threshold miscalibration in deepfake detection. Our code will be released.