The growing adoption of photorealistic 3D facial avatars, particularly those utilizing efficient 3D Gaussian Splatting representations, introduces new risks of online identity theft, especially in systems that rely on biometric authentication. While effective adversarial masking methods have been developed for 2D images, a significant gap remains in achieving robust, viewpoint-consistent identity protection for dynamic 3D avatars. To address this, we present AEGIS, the first privacy-preserving identity masking framework for 3D Gaussian avatars, concealing identity from automated face recognition while preserving the avatar's perceptual realism and animation usability. AEGIS applies adversarial perturbations to the base (DC) color coefficients of the Gaussian primitives, guided by an ensemble of pre-trained face verification networks, embedding the protection directly in the 3D representation so that it applies across viewpoints and animations, without retraining or modifying the avatar's geometry. Evaluated against black-box recognizers, AEGIS reduces face retrieval and verification accuracy to near-zero while maintaining high perceptual fidelity. It also preserves key facial attributes such as age, gender, and emotion, demonstrating strong privacy protection with minimal visual distortion. We further introduce a viewpoint-robust leakage metric that quantifies residual identity leakage across an avatar's rendered viewpoints, which single-view evaluation cannot capture.
Figures & tables
Figure 1: An adversary who obtains a 3D avatar can spoof RGB-based face verification, even under active liveness checks that prompt head rotation or expression changes and confirm identity from multiple viewpoints, because a controllable photorealistic avatar renders consistently from any viewpoint in real time. A 2D mask is applied to each rendered frame rather than embedded in the asset, so the raw avatar can be re-rendered from any viewpoint with no protection present. AEGIS instead embeds adversarial protection directly into the 3D representation, masking identity across viewpoints and poses so that the rendered avatar fails multi-view RGB verification
Figure 2: An overview of the AEGIS identity masking pipeline, which adversarially optimizes the DC color coefficients Ct of a 3D Gaussian avatar to suppress its identity under an ensemble of face recognizers F(⋅) . A reference embedding er is computed once from the original avatar G0 at the canonical camera vr and pose pr . Each PGD step samples camera parameters {vk}k=1K to cover diverse viewpoints, renders the current avatar Gt using R(⋅) , aligns the crops with A(⋅) , and extracts embeddings {ek} whose mean cosine similarity sˉ to er , averaged over the sampled viewpoints and the ensemble, defines the match / no-match logits ( sˉ and −sˉ ). A cross-entropy loss targeting “no match” is backpropagated to update Ct , yielding the masked avatar
Figure 3: Visualization of example identity masks obtained using AEGIS and reference 2D masking methods for three NeRSemble subjects
Verification (%) ↓
Retrieval (%) ↓
Viewpoint robustness ↓
Method
IRSE50
IR152
FaceNet
Mobile
Avg
Rank-1
Rank-50
VRL-AUC
L(τ⋆)
WC
Avatar renders
100.00
100.00
100.00
100.00
100.00
100.00
100.00
0.693±0.056
0.99±0.01
1.00±0.00
AEGIS RA ( ϵ=0.2 )
0.00
3.33
1.67
0.00
1.25±1.38
0.83±1.44
6.25±7.49
0.016±0.012
0.10±0.07
0.36±0.16
AEGIS base ( ϵ=0.1 )
18.33
51.67
26.67
11.67
27.08±15.16
24.58±11.27
61.25±10.23
0.105±0.032
0.59±0.10
0.98±0.01
AEGIS base ( ϵ=0.2 )
0.00
0.00
0.00
0.00
0.00
0.00
0.42±0.72
0.002±0.002
0.02±0.02
0.18±0.14
AEGIS base ( ϵ=0.3 )
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.000
0.00
0.03
Table 1: Privacy evaluation on the 60 NeRSemble subjects. Verification is the per-model rate at which the masked avatar still matches the real identity (TAR@FAR =10−3 , single reference); every scoring recognizer is unseen during masking. Retrieval (Rank-1/Rank-50) is pooled as mean ± std over the four black-box recognizers, against a gallery of 60 identity templates plus 10,177 CelebA distractors. Viewpoint robustness (VRL-AUC, leakage L(τ⋆) , worst-case WC) applies only to 3D representations; 2D methods have no persistent avatar. Lower is more private; bold marks the best per column and shading the recommended operating point
Method
SSIM ↑
PSNR (dB) ↑
FID ↓
Emotion (%) ↑
Gender (%) ↑
Race (%) ↑
Age (diff.) ↓
Avatar renders
1
∞
0
100
100
100
0
AEGIS RA ( ϵ=0.2 )
0.9347±0.0054
32.66±0.32
29.62±3.87
80.25
91.00
45.75
3.63
AEGIS base ( ϵ=0.1 )
0.9586±0.0037
35.03±0.23
22.52±2.59
87.50
91.25
57.25
2.87
AEGIS base ( ϵ=0.2 )
0.8837±0.0080
29.62±0.31
54.43±6.58
74.50
91.75
32.75
3.93
AEGIS base ( ϵ=0.3 )
0.8124±0.0116
26.78±0.40
75.96±7.71
73.25
92.00
25.75
5.05
LowKey ( ϵ=0.05 )
0.8528±0.0265
30.65±0.81
45.73±9.33
78.25
90.25
39.00
4.54
Table 2: Utility evaluation on the 60 NeRSemble subjects. SSIM, PSNR, and FID are reported as mean ± std; per-subject attribute-match rates (Emotion, Gender, Race) and the age difference are reported as means, as their per-subject standard deviations are dominated by the binary/quantized measurement structure. SSIM, PSNR, and age are pooled over subjects, while FID and attribute rates are computed over the evaluation folds. Bold marks the best value per column (excluding the unmasked reference) and shading marks the recommended operating point. Methods with strong attribute retention but no privacy (e.g., PixelDP ϵ=20 , cf. Tab. 1 ) leave the identity essentially intact
Figure 4: Privacy–quality trade-off for AEGIS variants, shown for a single example subject. Each point plots a variant’s mean cosine similarity to the genuine reference over the 81 rendered viewpoints of the VRL grid (x-axis, ↓ more private) against its SSIM (y-axis, ↑ higher quality), scored by the held-out IR152 recognizer ( τ⋆=0.256 ). The dashed line marks the verification threshold, separating de-identified (red, left) from verifiable (green, right) points. AEGIS RA achieves the best balance
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Temporal and privacy-preservation consistency of AEGIS ( ϵ=0.2 ) masking during motion, spanning expression and viewpoint changes, scored by the held-out IR152 recognizer
Category
Transformation
base (%) ↓
RA (%) ↓
Standard degradations
JPEG quality 70
0
3.33
JPEG quality 50
0
3.33
JPEG quality 30
0
3.33
JPEG quality 10
1.67
5.00
Iterative JPEG (5 × )
0
3.33
Iterative JPEG (10 × )
0
3.33
Appendix
Table 3: Verification rate (%) under image transformations for AEGIS base and RA ( ϵ=0.2 , 60 subjects, scored by the held-out IR152 recognizer at TAR@FAR =10−3 ). Lower is better ( ↓ )
Config
Description
Eyes
Nose
Lips
Skin
Forehead
Ears
Scalp
00 Uniform
No adaptation
1.0
1.0
1.0
1.0
1.0
1.0
1.0
01 Final RA*
Identity boost, smooth reduction
1.3
1.2
1.0
0.7
0.6
0.7
0.4
02 Boost Eye+Nose
Stronger boost on top-2 regions
1.5
1.4
1.0
0.7
0.6
0.7
0.4
03 Boost E+N+Lips
Include lips as identity-critical
1.5
1.4
1.3
0.7
0.6
0.7
0.4
04 Two-tier
Binary high/low allocation
1.4
1.4
1.4
0.4
0.4
0.4
0.4
05 Tight Smooth
Aggressive smooth-region reduction
1.3
1.2
1.0
0.5
0.4
0.5
0.2
Appendix
Table 4: Per-region ϵ multiplier configurations evaluated in the sweep. Region groups: Identity-critical (eyes, nose), Balanced (lips), Smooth (skin, forehead, ears, scalp). Unspecified regions inherit the base multiplier of 1.0
Config
Verification Rate (%) ↓
SSIM ↑
PSNR (dB) ↑
Avg. Sim. ↓
00 Uniform
0
0.899
31.05
−0.259
01 Final RA
0
0.937
33.74
−0.061
02 Boost Eye+Nose
20
0.950
34.99
+0.167
03 Boost E+N+Lips
0
0.949
34.95
+0.134
04 Two-tier
0
0.953
35.26
+0.182
05 Tight Smooth
30
0.960
36.14
+0.207
Appendix
Table 5: Quantitative results for each multiplier configuration, evaluated on 10 NeRSemble subjects (preliminary). Verification rate is the fraction of subjects verified above the TAR@FAR threshold ( τ=0.256 ). Lower Avg. Sim. provides a larger safety margin from the threshold
Method
Rank 1 (%) ↓
Rank 50 (%) ↓
Verification (%) ↓
SSIM ↑
PSNR (dB) ↑
FID ↓
Age (diff.) ↓
Race (%) ↑
Gender (%) ↑
Emotion (%) ↑
AEGIS (DC)
0
0
0
0.878
29.35
64.16
4.80
30
90
80
AC
0
0
0
0.584
20.65
141.37
8.80
10
90
30
Opacity
100
100
100
0.981
38.91
11.54
2.20
80
90
90
Position
0
0
0
0.711
23.59
109.24
9.40
30
90
40
Rotation
50
90
80
0.923
30.23
39.97
4.00
70
90
90
Scale
10
20
10
0.894
29.23
48.45
6.80
30
90
80
Appendix
Table 6: Quantitative evaluation of adversarial attacks applied to different components of the 3D Gaussian avatar representation, compared to AEGIS’s DC-based configuration (shaded). Preliminary study on 10 NeRSemble subjects under the ensemble surrogate protocol with IR152 held out. Lower is better for Rank- k /Verification/FID/Age; higher for SSIM/PSNR/Race/Gender/Emotion
Figure 6: Qualitative comparison of adversarial perturbations applied to different components of the 3D Gaussian avatar representation. Each column shows the masked render under an attack targeting a single component, alongside the original and AEGIS’s DC-based masking
Method
Rank 1 (%) ↓
Rank 50 (%) ↓
Verification (%) ↓
SSIM ↑
PSNR (dB) ↑
FID ↓
Age (diff.) ↓
Race (%) ↑
Gender (%) ↑
Emotion (%) ↑
AEGIS ( ℓ∞ -PGD)
0
0
0
0.878
29.35
64.16
4.80
30
90
80
DDN
80
100
90
0.977
38.36
12.39
4.00
40
90
100
ℓ∞ -FGSM ( ϵ=0.2 )
60
90
80
0.858
29.46
74.29
7.70
50
90
60
ℓ2 -PGD ( ϵ=6 )
70
90
80
0.975
37.76
12.94
4.50
50
90
90
ℓ2 -PGD ( ϵ=12 )
0
20
0
0.935
32.44
30.06
4.90
30
90
70
ℓ2 -FGSM ( ϵ=6 )
100
100
100
0.974
37.63
13.61
2.10
80
90
90
Appendix
Table 7: Quantitative comparison of adversarial optimization strategies (all targeting DC) under ℓ∞ and ℓ2 constraints, compared to AEGIS’s ℓ∞ -PGD configuration (shaded). Preliminary study on 10 NeRSemble subjects under the ensemble surrogate protocol with IR152 held out. Lower is better for Rank- k /Verification/FID/Age; higher for SSIM/PSNR/Race/Gender/Emotion
Figure 7: Qualitative comparison of identity masking under different adversarial optimization strategies (DDN, FGSM, and PGD under ℓ∞ and ℓ2 norms), all targeting DC and shown alongside the original render and AEGIS’s ℓ∞ -PGD masking
Figure 8: Extended qualitative visualization of avatar masking results for methods evaluated in the main text