Stop My Dancing! Understanding, Detecting and Attributing Motion-Aware Deepfake Videos

Authors: Fazhong Liu, Yan Meng, Tian Dong, Guoxing Chen, Haojin Zhu

Published: 2026-10-08 08:35:40+00:00

AI Summary

This paper addresses the emerging threat of Motion-Aware Deepfakes (MAD) generated by pose-guided diffusion models, which exhibit high global coherence. The authors construct the first MAD-specific benchmark dataset and propose MoDA, a novel defense framework. MoDA effectively detects and attributes MAD videos by combining spatial semantics with steganalysis-rich frequency features, demonstrating significant accuracy improvements and robustness against various attacks.

Abstract

Pose-guided diffusion models can now synthesize entire human figures in motion, spawning a new class of deepfakes: Motion Aware Deepfake (MAD) that have already reached hundreds of millions of viewers. To better understand this emerging threat, we construct the first MAD-specific benchmark and measurement framework, containing over 1.5 million frames that mix 1,363 real and 30,122 synthetic videos from six controllable generators, with realistic perturbations and open-world evaluation splits. Then, we dissect MAD and discover that, despite their global coherence, these videos betray faint yet reliable cues: because the model relies on limited input frames for motion synthesis, it must predict and simulate coherent movement at motion boundaries, thereby producing high-frequency artifacts along with model-specific spectral fingerprints. Based on the observations obtained from analysis on dataset, we propose MoDA, the first defense framework tailored to detect and attribute MAD videos. MoDA couples spatial semantics with steganalysis-rich frequency features via cross-domain alignment and multi-scale aggregation, achieving 94.8% in-distribution and 89.1% cross-dataset detection accuracy gains of 10% to 25% over prior work and 91.5% model attribution accuracy. MoDA achieves 81.94% accuracy on 200 clips produced by two unseen commercial MAD platforms, indicating promising zero-shot transfer, and 78.13% detection accuracy on 1,200 unseen MAD video clips (55k frames in total) collected from the open Internet. Under white-box, gray-box, and black-box adaptive attacks, MoDA maintains relatively stable detection and attribution performance while the accuracies of the baselines drop rapidly.

Key findings

MoDA achieved 94.8% in-distribution and 89.1% cross-dataset detection accuracy, outperforming prior work by 10-25%, and 91.5% model attribution accuracy. It also demonstrated promising zero-shot transfer with 81.94% accuracy on unseen commercial platforms and 78.13% on open-Internet MAD videos. MoDA maintained stable performance under white-box, gray-box, and black-box adaptive attacks, showcasing superior robustness.

Approach

MoDA leverages a dual-stream feature fusion mechanism that synergistically combines spatial-semantic and frequency cues. It uses a learnable high-pass filter to capture motion-induced high-frequency artifacts, a cross-domain feature alignment module to fuse these with spatial semantics, and a key-region focus module with attention to highlight suspicious motion boundaries. Multi-scale aggregation further enhances representational capacity for both detection and attribution.

Datasets

MAD dataset (custom-built, containing 1,363 real and 30,122 synthetic videos from six diffusion models), Open-Internet Split (1,200 clips from YouTube, TikTok, X), Commercial Split (200 clips from I2VControl and Wan2.2-Animate).

Model(s)

XCeption backbone (for spatial feature extraction), custom learnable high-pass filter, self-attention mechanisms, multi-layer perceptron (MLP) head for classification.

Author countries

China

← Previous