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

View on arXiv ← Back to list

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

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

AI Summary

This paper proposes D³ (Discrepancy Deepfake Detector), a novel framework for deepfake detection that improves generalization and robustness by training on multiple generators. D³ introduces a parallel network branch using a distorted image as a discrepancy signal to learn universal artifacts across various generators, leading to improved out-of-domain performance.

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

D³ achieves a 5.3% improvement in out-of-domain accuracy compared to state-of-the-art methods while maintaining in-domain performance. The method shows robustness against post-processing operations like Gaussian blur and JPEG compression. Ablation studies confirm the importance of all components of the proposed architecture.

Approach

D³ uses a two-branch network architecture. One branch processes the original image, and the other processes a distorted version (e.g., patch-shuffled) of the same image. A self-attention mechanism combines features from both branches, learning to identify universal artifacts that remain consistent despite the distortion. A final classifier predicts whether the image is real or fake.

Datasets

UniversalFakeDetect (UFD) and GenImage datasets, merged and scaled up to include 20 state-of-the-art generators (8 for training, 12 for testing).

Model(s)

Pre-trained CLIP:ViT-L/14 model as the feature extractor; a self-attention layer and a fully connected layer are trained.

Author countries

China, USA

← Previous