Work

VERITAS: Continual Learning for Robust AI-Generated Image Detection via Reconstruction Fingerprinting

Year: 2026
Type: article
Abstract: VERITAS (Continual Learning for Robust AI-Generated Image Detection via Reconstruction Fingerprinting) presents a research framework for detecting AI-generated images while addressing the challenge of catastrophic forgetting in continually evolving generative model environments. The proposed framework combines multi-strength diffusion reconstruction fingerprinting with continual learning techniques. Given an input image, a pretrained diffusion model reconstructs the image at multiple noise strengths. The resulting reconstruction residuals are stacked to form a fingerprint representation, which is subsequently classified as real or AI-generated using a convolutional neural network or frozen CLIP features. To support adaptation to newly introduced image generators, VERITAS incorporates continual learning strategies, including experience replay and Elastic Weight Consolidation (EWC). The framework evaluates detection performance across sequential generator domains, previously observed generators, and generators held out during training. Experiments conducted using curated subsets of GenImage, Artifact, and Synthbuster report a detection accuracy of up to 92.4% and a ROC-AUC of 0.972. In the reported stream-wise continual-learning experiments, experience replay reduces the forgetting measure from 0.392 to 0.050 and improves final average accuracy from 68.7% to 95.3% compared with naive fine-tuning. This work investigates the potential of reconstruction-based image fingerprints combined with continual learning to improve the robustness, adaptability, and generalization of AI-generated image detection systems. Keywords: AI-generated image detection, continual learning, diffusion models, reconstruction fingerprinting, catastrophic forgetting, image forensics, deep learning, CLIP, experience replay, Elastic Weight Consolidation.
Language: English
Open Access status: green
OpenAlex ID: W7235748982
FWCI:
Cites:
Cited by:
Related to: 10