Talking Face Generation (TFG) has made remarkable progress in synthesizing realistic speaker videos, evolving into two paradigms based on identity incorporation: image-based 2D TFG and video-based 3D TFG. Despite enabling efficient content creation, TFG also raises risks of unauthorized portrait forgery and privacy abuse, motivating proactive defenses. Existing defenses mainly target image-based 2D TFG, while, to our knowledge, no proactive defense has been developed for video-based 3D TFG. A straightforward solution is to apply existing image-based defenses independently to each video frame. However, experiments show that this adaptation is ineffective and computationally expensive: video-based 3D TFG lacks the image-reference modules and reverse diffusion processes exploited by existing defenses, and per-frame optimization incurs substantial overhead. To address this gap, we analyze the vulnerabilities of video-based 3D TFG and propose ErodedFace, an efficient proactive defense framework targeting reference-video preprocessing. ErodedFace uses a curriculum-inspired two-stage strategy: Stage I establishes scale-robust disruption through similarity-guided parameter inheritance and adaptive multi-scale optimization; Stage II improves perceptual fidelity and transformation robustness without weakening protection. Extensive experiments show that ErodedFace effectively protects portrait videos against unauthorized video-based 3D TFG, achieves a 10.0x speedup in defense generation over existing image-based 2D TFG defenses, and remains robust to common transformations and recent purification methods.
Figures & tables
Figure 1: ErodedFace is, to the best of our knowledge, the first proactive defense designed for unauthorized video-based 3D TFG. Existing proactive defenses mainly target image-based 2D TFG, where the protected reference image is directly consumed during generation. In contrast, video-based 3D TFG follows a fundamentally different paradigm, as shown in (a) and (b).
Figure 2: The basic workflow of the video-based 3D TFG method for acquiring personalized information.
Figure 3: Overview of ErodedFace. Stage I initializes the current-frame perturbation parameters through similarity-guided parameter inheritance, applies Fourier amplitude/phase modulation with a learnable spatial gate, and optimizes parsing erosion over the active unresolved-scale set. Stage II refines the protected frame through raw and differentiable-JPEG branches using perceptual and temporal losses, while an active parsing guard preserves the target erosion rate.
Method
SSIM ↑
PSNR ↑
FID ↓
Sync-C ↑
M-LMD ↓
Time(h)
AdvDM(-) ( Liang et al. 2023 )
0.4793
27.48
63.40
7.8606
2.6537
22.3
PhotoGuard ( Salman et al. 2023 )
0.5040
26.54
107.53
7.6222
3.1085
18.6
SDS(+) ( Xue et al. 2024 )
0.5453
27.57
127.02
7.9264
2.7472
12.7
SDS(-) ( Xue et al. 2024 )
0.5477
28.55
67.31
7.8797
2.6503
12.8
SDST(-) ( Xue et al. 2024 )
0.5530
27.55
100.80
7.7968
2.9659
19.5
Silencer-I ( Gan et al. 2025 )
0.5311
28.48
78.29
7.8652
2.6416
18.8
Table 1: Quantitative Comparisons with representative methods on protected videos. Higher SSIM, PSNR, and Sync-C and lower FID and M-LMD indicate better fidelity and audio-visual consistency. Best and second-best results are shown in bold and underlined, respectively. Time is reported for protecting a 45-second video.
Method
SSIM ↓
PSNR ↓
FID ↑
Sync-C ↓
M-LMD ↑
AdvDM(-) ( Liang et al. 2023 )
0.8999
29.58
21.15
6.7968
9.725
PhotoGuard ( Salman et al. 2023 )
0.8890
27.58
39.67
6.0465
16.70
SDS(+) ( Xue et al. 2024 )
0.8864
29.01
61.50
6.8855
9.032
SDS(-) ( Xue et al. 2024 )
0.8973
28.49
19.57
6.7258
12.06
SDST(-) ( Xue et al. 2024 )
0.8309
25.68
61.29
6.4674
13.15
Silencer-I ( Gan et al. 2025 )
0.8983
29.59
30.36
6.7450
10.02
Table 2: Quantitative comparison of downstream protection performance on videos generated by SyncTalk from protected reference videos. Lower SSIM, PSNR, and Sync-C and higher FID and M-LMD indicate stronger protection.
Method
JPEG
Resizing
PSNR ↓ /FID ↑ /Sync-C ↓ /M-LMD ↑
PSNR ↓ /FID ↑ /Sync-C ↓ /M-LMD ↑
AdvDM(-) ( Liang et al. 2023 )
29.76/17.70/6.9259/9.843
30.15/24.16/6.6624/10.26
PhotoGuard ( Salman et al. 2023 )
27.14 /71.36/ 6.0550 / 24.61
28.23/33.40/ 6.1544 / 12.45
SDS(+) ( Xue et al. 2024 )
29.04/55.74/6.5871/9.069
29.30/56.96/6.7124/8.859
SDS(-) ( Xue et al. 2024 )
27.25/62.46/7.1242/22.70
29.28/21.69/6.6871/10.18
SDST(-) ( Xue et al. 2024 )
27.97/32.16/6.1517/11.88
27.95 /31.88/6.3840/12.00
Table 3: Quantitative comparison of downstream protection performance after JPEG compression and resizing. The transformed protected videos are used as references for person-specific SyncTalk training. JPEG compression is performed with a quality factor of 75, while resizing downsamples each frame from 512 × 512 to 400 × 400 and subsequently restores it to 512 × 512. The two transformations are evaluated separately. Lower PSNR and Sync-C and higher FID and M-LMD indicate stronger protection.
Figure 4: Qualitative evaluation under JPEG compression and resizing. From top to bottom, we show the original reference frame and SyncTalk results obtained using ErodedFace-protected videos without transformation, after JPEG compression, and after resizing, respectively. Facial artifacts and structural distortions persist after both transformations, indicating that the protective effect is retained.
Stage-II
Mask
VGG
Param. Inherit.
Protected Image
SSIM ↑
PSNR ↑
FID ↓
Time (min)
–
✓
✓
✓
0.9613
34.97
21.08
12.91
✓
–
✓
✓
0.9549
29.10
15.24
117.5
✓
✓
–
✓
0.9038
33.12
55.90
25.73
✓
✓
✓
–
0.9285
34.13
24.52
297.5
✓
✓
✓
✓
0.9744
38.69
15.65
76.18
Table 4: Ablation study of ErodedFace. – denotes that the corresponding component is removed, while ✓ denotes that it is enabled. The last row represents the full configuration.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Method
SSIM ↓
PSNR ↓
FID ↑
Sync-C ↓
M-LMD ↑
DEGSTalk ( Deng et al. 2025 )
0.6984
18.96
46.62
1.0610
36.84
TalkingGaussian ( Li et al. 2024 )
0.6859
18.62
139.2
1.0995
47.95
InsTaG ( Li et al. 2025 )
0.7913
17.11
159.4
1.5457
61.21
Ground Truth
1.0000
∞
0.000
8.0789
0.000
Appendix
Table 5: Cross-model transferability of ErodedFace across video-based 3D TFG methods. The same ErodedFace protected reference videos are used to personalize each target model without re-optimization or model-specific adaptation, and all models are rendered using the same driving audio. Metrics are evaluated on the resulting generated videos. Lower SSIM, PSNR, and Sync-C and higher FID and M-LMD indicate stronger protection.
Figure 5: Qualitative cross model transferability of ErodedFace. Generated frames from SyncTalk, InsTaG, TalkingGaussian, and DEGSTalk exhibit pronounced facial corruption and structural inconsistencies, demonstrating that ErodedFace remains effective across different NeRF- and 3DGS-based TFG architectures.
Figure 6: Qualitative comparison with additional image-level defense methods on SyncTalk. For each method, we show the protected reference frames and the corresponding SyncTalk-generated frames. Existing image-level defenses still permit recognizable and structurally plausible synthesis, whereas ErodedFace induces pronounced facial corruption and structural distortions.
Method
SSIM ↓
PSNR ↓
FID ↑
Sync-C ↓
M-LMD ↑
AdvDM(+) ( Xue et al. 2024 )
0.8820
28.53
57.84
6.7685
9.552
Mist ( Liang and Wu 2023 )
0.8949
28.53
40.41
6.0717
13.91
FaceLock ( Wang et al. 2025 )
0.9048
29.85
16.14
7.9346
9.897
FaceShield ( Jeong et al. 2025 )
0.9130
29.68
10.37
7.8533
8.984
ErodedFace(ours)
0.8052
16.27
203.9
0.4276
176.8
Ground Truth
1.00
∞
0.00
8.1076
0.0000
Appendix
Table 6: Additional comparison with image-level defense methods on SyncTalk-generated videos. Each method is applied to the same reference videos, and the resulting protected videos are used for person-specific SyncTalk training under identical settings. Lower SSIM, PSNR, and Sync-C and higher FID and M-LMD indicate stronger protection.
Method
Diffpure
Freqpure
PSNR ↓ /FID ↑ /ID-S ↓ /M-LMD ↑
PSNR ↓ /FID ↑ /ID-S ↓ /M-LMD ↑
AdvDM(-) ( Liang et al. 2023 )
29.87/17.19/0.6382/10.78
28.94/ 22.51 / 0.5499 /12.71
AdvDM(+) ( Xue et al. 2024 )
29.34/16.79/0.6700/11.79
28.75/20.96/0.6043/10.69
PhotoGuard ( Salman et al. 2023 )
28.59 /16.50/0.6605/9.629
27.97 /19.32/0.5708/12.09
Mist ( Liang and Wu 2023 )
28.73/16.45/0.6404/11.86
28.20/19.61/0.5623/ 13.49
SDS(+) ( Xue et al. 2024 )
29.00/ 17.85 /0.6494/ 12.47
29.76/20.34/0.5886/9.551
Appendix
Table 7: Robustness to diffusion-based purification. Videos protected by each method are processed using DiffPure or FreqPure and subsequently used for person-specific SyncTalk training. Metrics are evaluated on the resulting generated videos. Lower PSNR and ID-S and higher FID and M-LMD indicate stronger residual protection after purification.
Figure 7: Qualitative results under diffusion-based purification. From top to bottom, we show the purified ErodedFace-protected frame and the corresponding SyncTalk results obtained after DiffPure and FreqPure purification.
Figure 8: Effect of the perturbation budget on facial landmark extraction. Although the landmark module is not explicitly targeted during optimization, its predictions become increasingly unstable as the perturbation budget grows. Clear landmark deviations emerge at ϵ=0.15 and become more pronounced under larger budgets.
Reference
SSIM ↓
PSNR ↓
FID ↑
Sync-C ↓
Clean
0.9320
29.65
9.872
8.09
Protected
0.6733
15.58
172.0
0.367
Appendix
Table 8: Generalization to additional portrait videos. We evaluate SyncTalk results on 40 additional portrait videos from HDTF. The clean setting uses the original reference videos, while the protected setting uses ErodedFace protected reference videos. The clear degradation from clean video to protected video indicates that ErodedFace remains effective on the additional HDTF videos.
Figure 9: Robustness of frequency-domain optimization to H.264 compression. Bars report macro-averaged rates over 11 identities, points denote identity-level temporal means, and error bars indicate actual confidence level bootstrap confidence intervals
The rapid development of generative portrait models has raised growing concerns about privacy leakage and identity misuse. In particular, audio-driven 3D talking face generation can reconstruct a reusable 3D portrait of a target person from a monocular video and animate it with arbitrary speech, making realistic identity impersonation alarmingly practical. Existing proactive defenses mainly operate in the visual domain by injecting subtle perturbations into acial regions to disrupt identity acquisition. However, such perturbations often compromise visual quality due to the strong structural priors and social sensitivity of human faces, and are easily weakened by common real-world transformations such as resizing. To overcome these limitations, we propose an imperceptible audio defense for audio-driven 3D talking face generation by shifting protection from the visual modality to the audio modality. Specifically,we exploit psychoacoustic masking to hide protective perturbations within perceptually masked frequency regions of the speech signal, thereby reducing perceptual distortion while suppressing reliable facial animation. Extensive experiments demonstrate that the proposed method effectively degrades 3D talking face generation while preserving favorable perceptual quality. These findings highlight psychoacoustically guided audio perturbations as a practical and promising direction for privacy-preserving portrait protection.
Rui-Qing Sun, Chen-Hao Cui, Hui-Yang Zhao +3
Beijing Institute of Technology Haidian Qu, Beijing Shi, China
Deformable 3D Gaussian Splatting (3DGS) has emerged as a popular method for real-time talking head synthesis, offering high-quality renderings at low latency. Tri-planes are a common choice for encoding Gaussians prior to deformation since they provide a compact and continuous representation. However, tri-plane encodings are limited by grid resolution and approximation errors introduced by projecting 3D volumetric fields onto 2D subspaces. Recent work has demonstrated the effectiveness of per-Gaussian embeddings for driving temporal deformations in 4D scene reconstruction. We introduce EmbedTalk, which leverages these embeddings to model speech-driven facial deformations for talking head synthesis. Comprehensive experiments show that EmbedTalk improves rendering quality, lip synchronisation, and motion consistency over previous 3DGS-based methods, while remaining competitive with state-of-the-art generative models. Replacing tri-plane features with embeddings also yields significantly more compact models that achieve 60+ FPS on a laptop GPU (RTX 2060 6 GB). Code and further details are available at https://arpita2512.github.io/EmbedTalk/
Arpita Saggar, Jonathan C. Darling, Duygu Sarikaya +1
School of Computer Science, University of Leeds · Leeds Institute of Medical Education, School of Medicine, University of Leeds
Diffusion-based audio-driven talking-head generation enables realistic portrait animation, but also introduces risks of misuse, such as fraud and misinformation. Existing protection methods are largely limited to a single modality, and neither image-only nor audio-only attacks can effectively suppress speech-driven facial dynamics. To address this gap, we propose SyncBreaker, a stage-aware multimodal protection framework that jointly perturbs portrait and audio inputs under modality-specific perceptual constraints. Our key contributions are twofold. First, for the image stream, we introduce nullifying supervision with Multi-Interval Sampling (MIS) across diffusion stages to steer the generation toward the static reference portrait by aggregating guidance from multiple denoising intervals. Second, for the audio stream, we propose Cross-Attention Fooling (CAF), which suppresses interval-specific audio-conditioned cross-attention responses. Both streams are optimized independently and combined at inference time to enable flexible deployment. We evaluate SyncBreaker in a white-box proactive protection setting. Extensive experiments demonstrate that SyncBreaker more effectively degrades lip synchronization and facial dynamics than strong single-modality baselines, while preserving input perceptual quality and remaining robust under purification. Code: https://github.com/kitty384/SyncBreaker.