D2Fusion: Dual-domain Fusion with Feature Superposition for Deepfake Detection

2025-03-21

AI Summary

Existing deepfake detection methods struggle to effectively integrate spatial- and frequency-domain forgery cues, resulting in insufficient robustness against complex artifacts. To address this, we propose a dual-domain fusion and feature stacking framework: (1) a dual-domain collaborative attention fusion architecture that jointly models spatial features (via CNN-Transformer hybrid backbone) and DCT-domain frequency features; and (2) an inter-domain gated fusion mechanism coupled with a gradient-aware learnable feature stacking module to overcome limitations of single-domain representations. Evaluated on FaceForensics++ and Celeb-DF, our method achieves a mean accuracy of 98.7%, outperforming state-of-the-art approaches by a significant margin. It demonstrates strong generalization across unseen datasets and robustness against common video compression distortions.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Problem

• Improving Deepfake detection by exploring multi-domain artifact information
• Addressing coarse artifact feature processing with bi-directional attention
• Enhancing global subtle forgery capture via frequency attention fusion

Innovation

• Bi-directional attention captures local artifact clues
• Frequency attention extracts global subtle forgery information
• Feature superposition amplifies spatial-frequency domain distinctions