Multi-Frequency Fusion for Robust Video Face Forgery Detection

2026-03-05 · Source: Apple Machine Learning Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Advanced, quick

Summary

Researchers have developed two novel face video forgery detectors, LFWS and LFWL, which achieve high accuracy with significantly smaller model sizes compared to existing wide or dual-stream backbone methods. Both detectors are built upon the Xception baseline model, which has 21.9 million parameters. LFWS integrates a low-frequency Wavelet-Denoised Feature (WDF) with a phase-only Spatial-Phase Shallow Learning (SPSL) map, while LFWL merges WDF with Local Binary Patterns (LBP). This lightweight fusion is achieved through an additional 1x1 convolution module, adding only 292 parameters and maintaining the total parameter count at 21.9 million, demonstrating improved efficiency.

Key takeaway

For AI scientists developing real-time deepfake detection systems, consider integrating lightweight fusion modules that combine handcrafted features like Wavelet-Denoised Features (WDF) with Spatial-Phase Shallow Learning (SPSL) or Local Binary Patterns (LBP). This approach can yield higher accuracy with minimal parameter overhead, allowing your models to run more efficiently on resource-constrained platforms without sacrificing detection performance.

Key insights

Lightweight fusion of handcrafted cues can significantly improve face forgery detection accuracy with minimal model overhead.

Principles

• Simpler models can outperform complex ones.
• Handcrafted features enhance deep learning models.

Method

Combine Wavelet-Denoised Features (WDF) with either Spatial-Phase Shallow Learning (SPSL) or Local Binary Patterns (LBP) using a 1x1 convolution.

In practice

• Integrate WDF and SPSL for forgery detection.
• Utilize LBP with WDF for alternative detection.

Topics

• Face Forgery Detection
• Lightweight Models
• Handcrafted Features
• Xception Architecture
• Wavelet Denoising

Best for: AI Scientist, Research Scientist, AI Researcher, Deep Learning Engineer, Computer Vision Engineer

Related on AIssential

• [P] VeridisQuo - open-source deepfake detector that combines spatial + frequency analysis and shows you where the face was manipulated — Combining spatial and frequency domain analysis enhances deepfake detection, especially for high-quality fakes.
• Depth-Semantic Alignment and Affinity-Guided Fusion for Structured Radar Point Cloud Generation — Fusing vision and radar data improves sparse, noisy millimeter-wave radar point clouds for perception tasks.
• Generalized Disguise Makeup Presentation Attack Detection Using an Attention-Guided Patch-Based Framework — A two-phase, attention-guided, patch-based framework effectively detects realistic disguise makeup attacks.
• M3D-Net: Multi-Modal 3D Facial Feature Reconstruction Network for Deepfake Detection — M3D-Net enhances deepfake detection by fusing RGB and 3D-reconstructed facial features.
• Deep Residual Injection for Full-Spectrum Forensic Signal Perception in Multimodal Large Language Models — MLLMs can achieve full-spectrum forensic signal perception by integrating low-level artifact signals via residual injection without losing semantic understanding.
• Detecting Deepfakes with Multivariate Soft Blending and CLIP-based Image-Text Alignment — Multimodal image-text alignment and blended data augmentation significantly improve deepfake detection generalization and robustness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Apple Machine Learning Research.