D2Fusion: Dual-domain fusion with feature superposition for Deepfake detection

Xueqi Qiu, Xingyu Miao, Fan Wan, Haoran Duan, Tejal Shah, Varun Ojha, Yang Long, Rajiv Ranjan

Abstract

As deepfake technology advances, distinguishing between authentic and forged media becomes increasingly difficult. Existing detection methods often focus on either spatial or frequency artifacts, leading to poor generalization across diverse manipulation techniques. This paper introduces D²Fusion, a framework that leverages dual-domain information through a unique feature superposition strategy. By treating image features as wave-like tokens, the model utilizes constructive and destructive interference to amplify forgery signals. D²Fusion demonstrates superior performance across multiple benchmarks, including FaceForensics++ and Celeb-DF, particularly in cross-dataset and multi-source detection scenarios.

1. Introduction

The rapid evolution of AI-generated content has made deepfakes a significant threat to digital integrity. While early detection methods could rely on obvious visual glitches, modern forgeries are often indistinguishable to the human eye. The core challenge in the field is no longer just detecting a specific type of fake, but developing models that generalize across various manipulation methods and environmental conditions.

2. The Generalization Challenge

Current deepfake detection research is broadly categorized into methods focusing on visual artifacts, object detection, and deep learning-based feature extraction. As shown in the comparison of existing methods, many current approaches suffer from over-reliance on specific domain artifacts or limited generalization capabilities.

Summary of existing deepfake detection methods and their limitations.

Most models fall into two traps: focusing exclusively on spatial artifacts (like blending boundaries) or frequency artifacts (like abnormal textures). D²Fusion addresses this by integrating both domains, moving beyond "coarse" feature processing to capture the subtle interactions between local organs and global textures.

3. Methodology: The D²Fusion Framework

The D²Fusion architecture employs a three-pronged strategy to refine feature extraction and fusion:

3.1 Bi-directional Attention Module (Spatial Domain)

To capture local positional information, this module utilizes horizontal and vertical average pooling. This allows the model to precisely localize forgery indicators, such as inconsistencies around the eyes or mouth, which are common in face-swapping and reenactment.

3.2 Fine-grained Spectral Attention Module (Frequency Domain)

Global clues like unnatural skin textures are often better captured in the frequency domain. D²Fusion uses a Discrete Cosine Transform (DCT) to partition features into multi-spectral components. This preserves high-frequency information that standard convolutional layers often discard, which is crucial for identifying high-quality forgeries.

3.3 Feature Superposition Strategy

The "secret sauce" of D²Fusion is its feature superposition strategy. Inspired by signal processing, the authors treat features as wave-like tokens. Using Euler’s formula, each token is assigned an amplitude (representing the feature value) and a phase (representing positional or semantic information).

  • Constructive Interference: Tokens with similar attributes (e.g., authentic features) have similar phases, amplifying their signal.
  • Destructive Interference: Tokens with differing attributes (e.g., authentic vs. artifact) undergo destructive interference, sharpening the boundary between real and fake content for the classifier.

4. Experimental Results

D²Fusion was evaluated on five major benchmarks: FaceForensics++ (FF++), Celeb-DF (v1 and v2), Deepfake Detection Challenge (DFDC), and DeeperForensics-1.0 (DFR). The model consistently outperformed state-of-the-art (SOTA) methods:

  • Intra-dataset Performance: On FF++, D²Fusion improved accuracy by nearly 2.5% over the baseline.
  • Cross-manipulation Robustness: When tested on unseen forgery types, the model achieved up to a 6.15% improvement in AUC.
  • Multi-source Detection: In complex scenarios with multiple forgery types, it surpassed SOTA methods by 14.62% in Precision and 10.26% in F1-score.
  • Cross-dataset Generalization: Training on FF++ and testing on DFR resulted in a 6.25% AUC improvement over the baseline, demonstrating its superior ability to handle real-world variations.

5. Conclusion and Future Work

D²Fusion represents a significant step forward in robust deepfake detection by effectively fusing spatial and frequency information through wave-like feature superposition. While the model shows exceptional performance on high-quality data, video compression remains a challenge, as it tends to "smear" high-frequency forgery clues. Future research will focus on more robust frequency transforms, such as wavelets, to maintain detection efficacy in highly compressed social media environments.

Summary updated

Abstract

Deepfake detection is crucial for curbing the harm it causes to society. However, current Deepfake detection methods fail to thoroughly explore artifact information across different domains due to insufficient intrinsic interactions. These interactions refer to the fusion and coordination after feature extraction processes across different domains, which are crucial for recognizing complex forgery clues. Focusing on more generalized Deepfake detection, in this work, we introduce a novel bi-directional attention module to capture the local positional information of artifact clues from the spatial domain. This enables accurate artifact localization, thus addressing the coarse processing with artifact features. To further address the limitation that the proposed bi-directional attention module may not well capture global subtle forgery information in the artifact feature (e.g., textures or edges), we employ a fine-grained frequency attention module in the frequency domain. By doing so, we can obtain high-frequency information in the fine-grained features, which contains the global and subtle forgery information. Although these features from the diverse domains can be effectively and independently improved, fusing them directly does not effectively improve the detection performance. Therefore, we propose a feature superposition strategy that complements information from spatial and frequency domains. This strategy turns the feature components into the form of wave-like tokens, which are updated based on their phase, such that the distinctions between authentic and artifact features can be amplified. Our method demonstrates significant improvements over state-of-the-art (SOTA) methods on five public Deepfake datasets in capturing abnormalities across different manipulated operations and real-life. Specifically, in intra-dataset evaluations, D2<math><msup is="true"><mrow is="true"><mi mathvariant="normal" is="true">D</mi></mrow><mrow is="true"><mn is="true">2</mn></mrow></msup></math>Fusion surpasses the baseline accuracy by nearly 2.5%. In cross-manipulation evaluations, it exceeds the baseline AUC by up to 6.15%. In multi-source manipulation evaluations, it exceeds the SOTA methods by up to 14.62% in P<math><mi is="true">P</mi></math>-value, 10.26% in F1<math><mrow is="true"><mi is="true">F</mi><mn is="true">1</mn></mrow></math>-score and 15.13% in R<math><mi is="true">R</mi></math>-value. In cross-dataset experiments, it exceeds the baseline AUC by up to 6.25%. For potential applications, D2<math><msup is="true"><mrow is="true"><mi mathvariant="normal" is="true">D</mi></mrow><mrow is="true"><mn is="true">2</mn></mrow></msup></math>Fusion can help improve content moderation on social media and aid forensic investigations by accurately identifying the tampered content.
BibTeX
@inproceedings{qiu2025d,
  author = {Xueqi Qiu and Xingyu Miao and Fan Wan and Haoran Duan and Tejal Shah and Varun Ojha and Yang Long and Rajiv Ranjan},
  title = {{D2Fusion: Dual-domain fusion with feature superposition for Deepfake detection}},
  booktitle = {Information},
  year = {2025},
  doi = {10.1016/j.inffus.2025.103087},
}
Figure