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

Fazhong Liu; Yan Meng; Tian Dong; Guoxing Chen; Haojin Zhu

cs.CVOct 8, 2026

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

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

Organizations: Shanghai Jiao Tong University Shanghai, China

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.

Figures & tables

Dataset Partition	Source (Model)	Input Format	Ref. Image Num	Motion Sources	Total Videos	Total Frames	Output Size
Real Videos	Real	/	/	1,363	1,363	65k	604 $\times$ 1080
Synthetic Videos	Mimicmotion ( Zhang et al., 2024 )	DWpose	40	340	9,350	448k	576 $\times$ 1024
	MusePose ( Tong et al., 2024 )	DWpose	40	340	9,350	448k	384 $\times$ 768
	Animate Anybody ( Hu, 2024 )	DWPose	40	340	9,350	448k	512 $\times$ 784
	Disco ( Wang et al., 2023a )	OpenPose	16	42	994	48k	256 $\times$ 256
	MagicAnimate ( Xu et al., 2024 )	DensePose	18	42	844	40k	512 $\times$ 512

Table 1. Details of MAD dataset. Total Videos denote 48-frame clips after filtering failed generations and low-quality outputs; counts therefore do not necessarily equal the Cartesian product of reference identities and motion sources.

Panel of Figure 2 — Figure 2. Multi-view evidence: understanding of MAD.

Model	$\mathbf{T}$ (Period)	$\mathbf{H}$ (Entropy)	$\mathbf{E_{low}/E_{high}}$ (Ratio)
Animate Anyone	Small	Low	Low
MimicMotion	Large	High	High
MusePose	Small	Medium-High	Highest

Table 4

Panel of Figure 3 — Figure 3. Model-specific fingerprints in frequency domain, demonstrating clear separability rooted in architecture.

Figure 5. Visual examples of Key-Region Focus Module. MoDA assigns higher attention weights to regions with salient motion boundaries.

Module	Filter	Alignment	Attention	Aggregation	End-to-end
Time (s)	0.026	0.062	0.081	0.014	0.183

Table 2. Runtime breakdown of MoDA (per 24 frames).

Category	Method	Original MAD-Dataset					Robustness under Post-processing
Category	Method	ACC	AUC	FPR	FNR	F1-Score	FS	LIC	PN	VE	Avg
Deepfake Detection	RECCE ( Cao et al., 2022 )	0.917	0.873	0.102	0.079	0.910	0.858	0.738	0.824	0.741	0.816
	DFGaze ( Xu et al., 2023 )	0.876	0.822	0.113	0.135	0.873	0.816	0.762	0.754	0.704	0.783
	XCeption ( Yang et al., 2023b )	0.886	0.834	0.085	0.131	0.889	0.956	0.828	0.509	0.724	0.781
	M2TR ( Wang et al., 2022 )	0.813	0.751	0.175	0.198	0.811	0.813	0.803	0.682	0.758	0.774
	FTCN ( Tan et al., 2024 )	0.671	0.599	0.346	0.312	0.669	0.624	0.592	0.467	0.517	0.574

Table 3. In-dataset evaluation results on MAD.

Training set	Method	Mimic	Muse	Anim	Avg
Mimic	RECCE ( Cao et al., 2022 )	0.968	0.864	0.771	0.868
	DFGaze ( Peng et al., 2024 )	0.897	0.731	0.256	0.628
	XCeption ( Rossler et al., 2019 )	0.733	0.664	0.570	0.656
	DeFake ( Sha et al., 2023 )	0.525	0.328	0.084	0.312
	MoDA	0.962	0.892	0.775	0.876
MusePose	RECCE ( Cao et al., 2022 )	0.774	0.956	0.449	0.728

Table 4. Cross-dataset evaluation results on MAD (recall on the MAD class).

Method	In-distribution		Cross-dataset
Method	ACC	AUC	ACC	AUC
Spatial	94.51	0.779	73.06	0.642
Frequency	91.63	0.688	84.53	0.752
Dual-Stream	95.80	0.756	82.16	0.724
Dual-Att-Stream	96.45	0.854	85.44	0.739
CDFA	94.91	0.764	83.32	0.758

Table 5. Ablation study of different model components.

Method	Mimic	MusePose	Animate	Avg
RECCE ( Cao et al., 2022 )	0.862	0.137	0.441	0.480
DFGaze ( Peng et al., 2024 )	0.896	0.707	0.920	0.841
XCeption ( Rossler et al., 2019 )	0.741	0.187	0.829	0.586
DeFake ( Sha et al., 2023 )	0.452	0.081	0.361	0.298
MoDA	0.997	0.858	0.890	0.915

Table 6. Model tracing evaluation results on MAD (ACC).

Variant	AUC	FPR	Memory	Latency
Full MoDA (platform)	0.969	0.023	1.6GB	7.6 ms/frame
Lightweight (mobile)	0.756	0.067	164MB	2.0 ms/frame

Table 7. Full MoDA vs. lightweight variant.

Figure 6. Examples from commercial MAD platforms.

Panel of Figure 7 — Figure 7. Detection under adaptive attacks and perturbation.

Defense	White-box PGD	Gray-box MI-FGSM
None	0.394	0.554
Adversarial training	0.523 (+18.2%)	0.703 (+14.9%)
Input preprocessing	0.417 (+3.3%)	0.619 (+6.5%)

Table 8. Mitigation results against adaptive attacks (ACC).

Explore similar work

Attribution-Guided Multimodal Deepfake Detection via Cross-Modal Forensic Fingerprints
Apr 29, 2026Wasim Ahmad, Wei Zhang, Xuerui MaoAudio Deepfake Detection Cross-Modal Representation Learning
School of Interdisciplinary Science, Beijing Institute of Technology, Beijing 100081, China
Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection
Aug 7, 2026Xuechao Zou, Shun Zhang, Kai Li +6Explainable Deepfake Detection Deepfake Detection
Beijing Jiaotong University · Tsinghua University · Ant Group
DF26: We Cannot Tell Fake From Real Anymore
Sep 7, 2026Severyn Shykula, Andrii Yermakov, Ivan Samarskyi +3Text-to-Video Generation OOD Generalization
Ukrainian Catholic University · Faculty of Electrical Engineering, Czech Technical University in Prague · Hover Inc., USA +1

Apr 29, 2026, cs.CVJ/K move · Enter open · S save

Attribution-Guided Multimodal Deepfake Detection via Cross-Modal Forensic Fingerprints

Wasim Ahmad, Wei Zhang, Xuerui Mao

School of Interdisciplinary Science, Beijing Institute of Technology, Beijing 100081, China

Audio Deepfake Detection Cross-Modal Representation Learning Digital Forensics Deepfake Detection

Open paper AskPDF arXiv

Audio-visual deepfakes have reached a level of realism that makes perceptual detection unreliable, threatening media integrity and biometric security. While multimodal detection has shown promise, most approaches are binary classification tasks that often latch onto dataset-specific artifacts rather than genuine generative traces. We argue that a detector incapable of identifying how a video was forged is likely learning the wrong signal. Unlike binary detection, attribution-guided learning imposes a stronger geometric constraint on the shared embedding space, forcing the model to encode generator-specific forensic content rather than shortcuts. We propose the Attribution-Guided Multimodal Deepfake Detection (AMDD) framework, which jointly learns to detect and attribute manipulation. AMDD treats generator attribution as a structured regularization that constrains representation geometry toward forensically meaningful features. We introduce a Cross-Modal Forensic Fingerprint Consistency (CMFFC) loss to enforce alignment between generator-induced artifacts in visual and audio streams. This exploits the fact that coherent manipulation leaves correlated traces across modalities, grounded in the physical coupling between speech and facial articulation that synthetic pipelines routinely disrupt. Architecturally, we pair a ResNet50 with temporal attention for visual encoding against a pretrained ResNet18 for mel spectrograms, closing the encoder capacity gap found in prior models. On FakeAVCeleb, AMDD achieves 99.7% balanced accuracy and 99.8% AUC with 95.9% attribution accuracy. Cross-dataset evaluation on DeepfakeTIMIT, DFDM, and LAV-DF confirms that real video detection generalizes robustly, while fake detection on unseen generators remains an open challenge that we analyze in depth.

Abstract

Explore similar work

Attribution-Guided Multimodal Deepfake Detection via Cross-Modal Forensic Fingerprints

Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection

DF26: We Cannot Tell Fake From Real Anymore