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

TL;DR

This work addresses the problem of universal deepfake detection across multiple generators by moving from the traditional train-on-one/test-on-many to a train-on-many/test-on-many paradigm. It introduces D$^3$, a discrepancy-based two-branch detector that processes an original image and a patch-shuffled distorted version, extracting shared artifacts with a CLIP-based backbone and self-attention to improve generator-invariant detection. Through large-scale experiments merging the UFD and GenImage datasets, the method demonstrates a 5.3 percentage-point improvement in OOD accuracy while maintaining strong in-domain performance, outperforming state-of-the-art methods on overall metrics. The results suggest that learning from discrepancy, combined with robust pre-trained representations, yields more generalizable deepfake detectors suitable for rapid evolution of generative models.

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

Figures (5)

• Figure 1: The overall framework of our Discrepancy Deepfake Detector (D$^\textbf{3}$). The original image and its patch-shuffled variant are inputted into the pre-trained vision model to extract two features. We then utilize the self-attention module to encourage the learning of shared artifacts between the features. Finally, we use a single-layer linear classifier to get the final prediction.
• Figure 1: Visualization of uniquely detected samples. We present a selection of samples from both in-domain and out-of-domain, that are accurately classified by our approach, yet erroneously classified by the UFD ufd, with a discrepancy in classification confidence exceeding 0.8.
• Figure 2: Mean accuracy of training on scaled-up datasets on the testing set. We design an experiment to show the performance of different methods when gradually adding generators into the training pool. To avoid occasionality, we add the generator in three distinct random orders and average the corresponding test results. (a) "In-domain" means 1 shared training generator (the first generator in any given order). (b) "Out-of-domain" means 12 shared OOD generators.
• Figure 2: Visualizations of fake image samples along with the corresponding occlusion maps occlusion for each detector. These images represent cases accurately identified by $D^3$ where UFD failed.
• Figure 3: Total mean accuracy results of robustness to post-processing operations on the testing set. We conducted experiments on Gaussian blurring ranging from 0 to 2 and JPEG compression ranging from 30 to 100 to verify the robustness of different methods.