OUT-OF-DISTRIBUTION AI-GENERATED IMAGE DETECTION WITH ZERO-PHASE DECONVOLUTION FOR RESIDUAL WHITENING
Abstract
Current AI-generated image detection methods struggle in out-of-distribution testing conditions because detectors are usually not robust when evaluated on generators not included in the training set. This work explores a way to mitigate the forensic bias between sources of images -whether real or generated -by removing generator-specific fingerprints. This new training process can be approximated by pre-processing involving a deconvolution operation designed to remove correlated noise patterns inducing loss of generalization. Indeed, we show that whitening the spatial distribution of the image noise residuals can significantly improve the performance of detectors on images coming from generators unseen during training. In particular, an end-to-end CNN model trained on whitened image residuals improves detection by 12.6% on average compared to a training involving original images.
Related
- Goodness-of-fit testing for the Pareto type-I distribution based on a mean residual life characterization
- Attribution-State Conformal Forecasting: Residual-Independent Explanation States for Uncertainty Calibration Under Distribution Shift
- Steer-to-Detect: Probing Hidden Representations for Detection of LLM-Generated Texts
- Robust Detection of LLM-Generated Text under Contamination
- Evaluating Xception with Multitask Self-Supervised Learning for AI-Generated Image and Deepfake Video Detection
BibTeX
@misc{pilavcı2026outofdistribution,
title = {{OUT-OF-DISTRIBUTION AI-GENERATED IMAGE DETECTION WITH ZERO-PHASE DECONVOLUTION FOR RESIDUAL WHITENING}},
author = {Yusuf Yigit Pilavcı and Antoine Mallet and Vincent Itier and Jan Butora and Jérémie Boulanger and Patrick Bas},
howpublished = {HAL},
year = {2026},
month = oct,
url = {https://hal.science/hal-05781026v1},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - OUT-OF-DISTRIBUTION AI-GENERATED IMAGE DETECTION WITH ZERO-PHASE DECONVOLUTION FOR RESIDUAL WHITENING
AU - Yusuf Yigit Pilavcı
AU - Antoine Mallet
AU - Vincent Itier
AU - Jan Butora
AU - Jérémie Boulanger
AU - Patrick Bas
PY - 2026
DA - 2026/10/08
PB - HAL
UR - https://hal.science/hal-05781026v1
AB - Current AI-generated image detection methods struggle in out-of-distribution testing conditions because detectors are usually not robust when evaluated on generators not included in the training set. This work explores a way to mitigate the forensic bias between sources of images -whether real or generated -by removing generator-specific fingerprints. This new training process can be approximated by pre-processing involving a deconvolution operation designed to remove correlated noise patterns inducing loss of generalization. Indeed, we show that whitening the spatial distribution of the image noise residuals can significantly improve the performance of detectors on images coming from generators unseen during training. In particular, an end-to-end CNN model trained on whitened image residuals improves detection by 12.6% on average compared to a training involving original images.
ER -