UniGADD: Universal Generator-Agnostic Deepfake Detector

Authors: Bernard Tomczyk, Agil Aghasanli, Plamen Angelov

Published: 2026-11-05 08:34:07+00:00

AI Summary

This paper introduces UniGADD, a Universal Generator-Agnostic Deepfake Detector, which uses a two-stage supervised learning approach to detect deepfakes. It achieves high accuracy on known deepfake generators while maintaining robust performance on unseen ones. The method refines the embedding space of real samples to enhance generalizability against novel deepfake generation techniques.

Abstract

In this paper, we introduce Universal Generator-Agnostic Deepfake Detector (UniGADD), a novel method that employs supervised learning to achieve high accuracy on known deepfake generators, while maintaining robust performance on previously unseen ones. The proposed approach follows a two-stage optimisation process. In the first stage, a contrastive loss encourages the model to learn discriminative feature embeddings from real and known fake images, resulting in strong performance within the training domain. In the second stage, the embedding space is refined by promoting inter-cluster separation and intra-cluster compactness, applied exclusively to real samples. This refinement enhances generalisability, enabling the method to exhibit improved robustness against unseen deepfake generation techniques. UniGADD achieves accuracy on par with state-of-the-art methods for known generators, while significantly outperforming them on unseen cases, demonstrating its scalability and practical applicability for adversarial content detection.


Key findings
UniGADD achieves state-of-the-art accuracy on known deepfake generators and significantly outperforms existing methods on unseen cases, with an overall unseen accuracy of 85.53% in the Leave-One-Out setting. It also demonstrates graceful degradation and consistent robustness in Leave-N-Out scenarios, even when up to 13 generators are excluded. The method's generalisation ability is robust across different Vision Transformer backbones.
Approach
UniGADD employs a two-stage optimization process. First, a supervised contrastive loss encourages the model to learn discriminative feature embeddings from real and known fake images. Second, the embedding space is refined by promoting inter-cluster separation and intra-cluster compactness, applied exclusively to real samples, which enhances generalisability to unseen deepfake generation techniques.
Datasets
AIGC datasets (Fan et al., 2024)
Model(s)
DINOv2 ViT-L/14 backbone (default), CLIP ViT-B/32, CLIP ViT-L/14, DINOv2-S/14, DINOv2-B/14
Author countries
United Kingdom