Work

Diffusion-generated face image identification technique based on diffusion reconstruction error feature

Year: 2025

Type: article

Abstract: Traditional deep learning detectors often struggle to generalise when detecting diffusion-generated content. To address this, we propose DIRE, a generalised detector leveraging reconstruction error image representation. The framework standardises facial feature spaces through constrained feature learning and introduces a gradient suppression algorithm to filter abnormal gradients, preventing shortcut learning and enhancing generalisable feature extraction. Experiments on hybrid datasets validate DIRE's effectiveness in four cross-domain tasks (O&C&M→I, O&C&I→M, O&I&M→C, and I&C&M→O). Ablation studies confirm the synergy of feature standardisation and gradient suppression, reducing bias by 97.6% and parameters by 42%, while accelerating inference by 2.3×. DIRE achieves 98.2% and 96.7% accuracy on two tasks (O&C&I→M and O&M&I→C), outperforming state-of-the-art methods by 5.3% while maintaining computational efficiency. This study advances generative face detection through dual optimisation, offering a lightweight framework for financial identity verification and social media content moderation.

Source: International Journal of Information and Communication Technology

DOI: 10.1504/ijict.2025.146364

Authors: Yunyue Peng, Yuan Liu, Tianqi Xie, Wei Zeng

Institutions: Chengdu University of Technology, University of Electronic Science and Technology of China, China Electric Power Research Institute

Language: English

Topic: Digital Media Forensic Detection

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Funders: National Natural Science Foundation of China, Natural Science Foundation of Sichuan Province

Awards: G2907891670, G4991683698