Research map / Image generation / Editing and style transfer
ToonerGAN: Reinforcing GANs for Obfuscating Automated Facial Indexing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ToonerGAN, a three-step GAN (style learning, style distillation with a StyleGAN2 decoder, de-identification) that turns face images into cartoon avatars to defeat automated face recognition while keeping them human-recognizable, and releases the ToonSet face-toon dataset.
Problem
Automated facial indexing and recognition threaten personal privacy, and classic de-identification (blurring, pixelation, GAN-based replacement) either fails against recognizers or destroys human recognizability.
Approach
Shift face images toward a cartoon distribution with an identity-encoder/style-decoder/de-identification-decoder generator, using a StyleGAN2-based high-resolution decoder guided by a pSp encoder and an identity loss (AdaFace cosine) so that the output is 1024x1024 toon avatars that evade face recognition yet stay recognizable to humans.
Key results
- On CelebA-HQ cartoon re-identification, ToonerGAN reaches Recall@1 of 0.015% and Recall@10 of 0.060%, versus 76.28% and 80.94% for the original images and 0.02%/0.52% for DeIDGAN.
- On LFW, ToonerGAN achieves TPR@0.001 FPR of 0.009 (lower is better), the best among compared methods (RiDDLE 0.016, original 0.965).
- Human study: 100% closed-set identification of original faces from toons and 92.35% toon-verification accuracy; MTCNN detection rate of 99.56%. Average inference time 68.69 ms on one A100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ToonerGAN instead of DeIDGAN | Face de-identification on CelebA-HQ · CelebA-HQ | Recall@k=1: 0.02% -> 0.015% | Helped |
| Use ToonerGAN instead of DeIDGAN | Face de-identification on CelebA-HQ · CelebA-HQ | Recall@k=10: 0.52% -> 0.060% | Helped |
| Use ToonerGAN instead of DeIDGAN | Face de-identification on CelebA-HQ · CelebA-HQ | Detection rate: 99.98% -> 99.56% | Hurt |
| Use ToonerGAN instead of RiDDLE | Face de-identification on LFW view 2 · LFW view 2 | TPR@0.001 FPR: 0.016 -> 0.009 | Helped |
| Use ToonerGAN instead of RiDDLE | Identity similarity on CelebA-HQ · CelebA-HQ | ID similarity: 0.056 -> 0.104 | Hurt |
| Add Steps 2+3 after Step 1 (ablation) | ToonerGAN · CelebA-HQ | Recall@k=1: 8.08 -> 0.015 | Helped |
| Add Steps 2+3 after Step 1 (ablation) | ToonerGAN · CelebA-HQ | Recall@k=10: 12.79 -> 0.06 | Helped |
| Add Steps 2+3 after Step 1 (ablation) | ToonerGAN · LFW view 2 | TPR@0.001 FPR: 0.19 -> 0.009 | Helped |
| Add Steps 2+3 after Step 1 (ablation) | ToonerGAN · CelebA-HQ | ID similarity: 0.355 -> 0.104 | Helped |
| Add Steps 2+3 after Step 1 (ablation) | ToonerGAN · CelebA-HQ | Detection rate: 99.23 -> 99.56 | Helped |
| Add Steps 2+3 after Step 1 (ablation) | ToonerGAN · ToonSet | ID similarity: 0.322 -> 0.1421 | Helped |
Surprising findings
- ToonerGAN's ID similarity (0.104) is not the lowest in Table 3; RiDDLE (0.056) and CIAGAN (0.068) score lower, so the paper claims only comparable identity similarity.
- Blurring and pixelation nearly eliminate re-identification but reduce face detection rate to 6.58% and 48.54%, and LIVE barely changes recognition results.
Limitations
- Toon avatars are generated by a commercial website (toonme.com), and the 23k pairs derive from that tool's style.
- Identity similarity against CurricularFace is not the best among compared methods (0.104 vs 0.056 for RiDDLE).
- Human attribute recognition is weaker for race (80.10%) and age (68.67%).
Experimental setup
Datasets: ToonSet FFHQ CelebA-HQ LFW CASIA-WebFace
Benchmarks: CelebA-HQ LFW
Metrics: Recall@k TPR@0.001 FPR ID similarity face detection rate human identification accuracy human verification accuracy human attribute accuracy
Baselines: Blurring Pixelation NEO ANet LIVE CIAGAN DeIDGAN Identity Transformer IVFG IdentityMask RiDDLE DualStyleGAN
Builds on: StyleGAN2 pSp encoder AnimeGAN2 DualStyleGAN AdaFace
Closest papers by meaning
- SwiftAvatar: Efficient Auto-Creation of Parameterized Stylized Character on Arbitrary Avatar Engines (AAAI 2023 · similarity 0.85)
- Attribute-Preserving Face Dataset Anonymization via Latent Code Optimization (CVPR 2023 · similarity 0.85)
- Interpreting the Latent Space of GANs for Semantic Face Editing (CVPR 2020 · similarity 0.85)
- Low-Rank Head Avatar Personalization with Registers (NeurIPS 2025 · similarity 0.83)
- SDGAN: Disentangling Semantic Manipulation for Facial Attribute Editing (AAAI 2024 · similarity 0.83)
- AvatarArtist: Open-Domain 4D Avatarization (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.