Unsupervised Deepfake Detection via Camera Source Clustering and Temporal-Spatial Features

View at publisher Summarise with VIRAT
Authors: Zhuang Li, Chenxu Lu, Zahid Akhtar, Tong Qiao
Cited by: 0
References: 80

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Artificial Intelligence (AI) is transforming diverse domains like cybersecurity, robotics, and healthcare but are also being misused for harmful purposes. A notable example is deep fakes that involve audio or visual content artificially generated or altered using deep neural networks. Current deepfake detection methods rely predominantly on supervised learning paradigm, which faces issues like poor interpretability and the impracticality of acquiring extensive labeled data. This paper proposes a novel unsupervised deepfake detection framework to overcome these limitations. By utilizing camera source clustering based on inherent noise patterns, the method decomposes large-scale deepfake clustering into smaller and more manageable problems. It further extracts and aggregates robust inter-frame and intra frame features to highlight temporal and spatial inconsistencies in deepfake videos. Extensive experiments validate that the proposed training-free framework achieves high detection accuracy and relevant generalization across multiple datasets, rivalling the training-based baselines. Moreover, the proposed scheme reduces reliance on labeled data, enhances interpretability, and paves the way for scalable and cost-effective deepfake detection solutions in real-world scenarios. Our source codes have been released at https://github.com/Q-bigbig/unsupervised-deepfake.

Fields of Study & Subjects

Digital Media Forensic Detection Image and Video Quality Assessment Generative Adversarial Networks and Image Synthesis

Publication details

Journal / source
IEEE Transactions on Multimedia
Volume
28
Pages
8093–8108
Publication date
1 January 2026
Language
EN