DATA: Multi-Disentanglement based Contrastive Learning for Open-World Semi-Supervised Deepfake Attribution

TL;DR

Open-world deepfake attribution faces overfitting to dataset biases and the challenge of unknown novel classes. The authors propose DATA, a multi-disentanglement framework that combines feature-level disentanglement through Deepfake Basis Exploration with Orthonormal Deepfake Bases and instance-level disentanglement via Augmented-Memory Clustering, guided by Bases and Center contrastive losses. The approach yields state-of-the-art results on OSS-DFA and OW-DFA++ benchmarks, particularly improving novel-class generalization and robustness to perturbations. By focusing on forgery-relevant, technique-agnostic features and steadily augmenting memory prototypes, DATA enhances generalization to unseen deepfake techniques with practical implications for real-world attribution tasks.

Abstract

Deepfake attribution (DFA) aims to perform multiclassification on different facial manipulation techniques, thereby mitigating the detrimental effects of forgery content on the social order and personal reputations. However, previous methods focus only on method-specific clues, which easily lead to overfitting, while overlooking the crucial role of common forgery features. Additionally, they struggle to distinguish between uncertain novel classes in more practical open-world scenarios. To address these issues, in this paper we propose an innovative multi-DisentAnglement based conTrastive leArning framework, DATA, to enhance the generalization ability on novel classes for the open-world semi-supervised deepfake attribution (OSS-DFA) task. Specifically, since all generation techniques can be abstracted into a similar architecture, DATA defines the concept of 'Orthonormal Deepfake Basis' for the first time and utilizes it to disentangle method-specific features, thereby reducing the overfitting on forgery-irrelevant information. Furthermore, an augmented-memory mechanism is designed to assist in novel class discovery and contrastive learning, which aims to obtain clear class boundaries for the novel classes through instance-level disentanglements. Additionally, to enhance the standardization and discrimination of features, DATA uses bases contrastive loss and center contrastive loss as auxiliaries for the aforementioned modules. Extensive experimental evaluations show that DATA achieves state-of-the-art performance on the OSS-DFA benchmark, e.g., there are notable accuracy improvements in 2.55% / 5.7% under different settings, compared with the existing methods.

Figures (8)

• Figure 1: The OSS-DFA benchmark consists of diverse known and novel deepfake classes. Meanwhile, the training set and test set share the same label space and our goal is to disentangle interwoven novel samples.
• Figure 2: An overview of the DATA framework. The DBE module is used to explore the 'orthonormal deepfake basis set' and disentangle method-specific feature $\mathbf{f}^s$. The AMC module is used to conduct the instance-level disentanglement with the assistance of an Augmented-Memory.
• Figure 3: An overall pipeline of the DBE module.
• Figure 4: A brief schematic of the AMC module.
• Figure 5: The hyper-parameters analysis of loss function.