D$^3$: Scaling Up Deepfake Detection by Learning from Discrepancy

Authors: Yongqi Yang, Zhihao Qian, Ye Zhu, Olga Russakovsky, Yu Wu

Published: 2024-04-06 10:45:02+00:00

Comment: 13 pages, 3 figures, accepted by CVPR 2025

AI Summary

This paper introduces D$^3$ (Discrepancy Deepfake Detector), a novel framework for universal deepfake detection designed for a scaled-up "train-on-many and test-on-many" scenario. It addresses the challenges of existing methods that either struggle to learn comprehensive artifacts from multiple generators or compromise in-domain performance for out-of-domain generalization. D$^3$ achieves superior generalization and robustness by deconstructing universal artifacts through a parallel network branch that leverages a distorted image feature as an extra discrepancy signal.

Abstract

The boom of Generative AI brings opportunities entangled with risks and concerns. Existing literature emphasizes the generalization capability of deepfake detection on unseen generators, significantly promoting the detector's ability to identify more universal artifacts. This work seeks a step toward a universal deepfake detection system with better generalization and robustness. We do so by first scaling up the existing detection task setup from the one-generator to multiple-generators in training, during which we disclose two challenges presented in prior methodological designs and demonstrate the divergence of detectors' performance. Specifically, we reveal that the current methods tailored for training on one specific generator either struggle to learn comprehensive artifacts from multiple generators or sacrifice their fitting ability for seen generators (i.e., In-Domain (ID) performance) to exchange the generalization for unseen generators (i.e., Out-Of-Domain (OOD) performance). To tackle the above challenges, we propose our Discrepancy Deepfake Detector (D$^3$) framework, whose core idea is to deconstruct the universal artifacts from multiple generators by introducing a parallel network branch that takes a distorted image feature as an extra discrepancy signal and supplement its original counterpart. Extensive scaled-up experiments demonstrate the effectiveness of D$^3$, achieving 5.3% accuracy improvement in the OOD testing compared to the current SOTA methods while maintaining the ID performance. The source code will be updated in our GitHub repository: https://github.com/BigAandSmallq/D3

Key findings

D$^3$ significantly improves Out-Of-Domain (OOD) testing accuracy by 5.3% compared to state-of-the-art methods while successfully maintaining In-Domain (ID) performance. It achieved the highest overall mean accuracy of 90.7% and an average precision (AP) of 96.8%. The method also demonstrated superior robustness against various post-processing operations like Gaussian blurring and JPEG compression.

Approach

The D$^3$ framework employs a dual-path design where both the original image and its patch-shuffled counterpart are inputted into a pre-trained vision model (CLIP:ViT-L/14) to extract distinct visual features. These features are then stacked and processed by a self-attention layer to learn shared artifact correlations between the original and distorted inputs. A final linear classifier uses these correlated features to make the real/fake prediction.

Datasets

UniversalFakeDetect, GenImage

Model(s)

CLIP:ViT-L/14 (as feature extractor), Self-Attention Layer, Single Fully Connected Layer

Author countries

China, USA

← Previous