Research map / Image generation / Editing and style transfer
Generative Unlearning for Any Identity
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 generative identity unlearning, removing a single source identity from a pre-trained 3D GAN (EG3D) using one image via an extrapolated target latent (UFO) and local, adjacency-aware, and global preservation losses (GUIDE).
Problem
Pre-trained GANs combined with GAN inversion can reconstruct and edit any person's face, even not in training data, raising privacy concerns; the goal is to make the generator unable to produce a specific identity given only one image while preserving overall generation quality.
Approach
Define the unlearning target as an extrapolation from the source latent away from the average latent (UFO), then fine-tune the generator to map the source latent to the target's rendering, adding adjacency-aware neighborhood losses and a global perceptual preservation loss on random latents.
Key results
- On the Random scenario, GUIDE reduces ID similarity from 0.19 (baseline) to 0.14 and FIDpre from 11.73 to 10.80.
- In the multi-image OOD test on CelebAHQ, GUIDE reduces FIDpre from 9.52 (baseline) to 7.88 and ∆FIDreal from 4.75 to 3.34, while IDothers is 0.17 vs 0.28 for baseline.
- Adding the global preservation loss reduces FIDpre from 9.52 to 4.63 in the ablation on CelebAHQ.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use an extrapolated target latent instead of the average latent (ablation) | GUIDE baseline · Random, In-Domain (FFHQ), and Out-of-Domain (CelebAHQ) | ID: 0.19 ± 0.09 -> 0.12 ± 0.06; 0.16 ± 0.07 -> 0.05 ± 0.06; 0.12 ± 0.06 -> 0.02 ± 0.05 | Helped |
| Use an extrapolated target latent instead of the average latent (ablation) | GUIDE baseline · Random, In-Domain (FFHQ), and Out-of-Domain (CelebAHQ) | FIDpre: 11.73 ± 2.74 -> 14.28 ± 3.34; 9.00 ± 1.15 -> 12.78 ± 1.82; 9.52 ± 1.53 -> 13.02 ± 3.20 | Hurt |
| Add adjacency-aware unlearning loss (ablation) | GUIDE with extrapolated wt · Out-of-Domain (CelebAHQ), multi-image test | IDothers: 0.15 ± 0.07 -> 0.14 ± 0.07 | Helped |
| Add adjacency-aware unlearning loss (ablation) | GUIDE with extrapolated wt · Out-of-Domain (CelebAHQ) | FIDpre: 13.02 ± 3.20 -> 13.63 ± 3.52 | Hurt |
| Add global preservation loss (ablation) | GUIDE with extrapolated wt and Ladj · Random, In-Domain (FFHQ), and Out-of-Domain (CelebAHQ) | FIDpre: 19.65 ± 4.90 -> 10.80 ± 2.70; 13.53 ± 2.08 -> 8.00 ± 1.20; 13.63 ± 3.52 -> 7.88 ± 1.96 | Helped |
| Add global preservation loss (ablation) | GUIDE with extrapolated wt and Ladj · Out-of-Domain (CelebAHQ), multi-image test | IDothers: 0.14 ± 0.07 -> 0.17 ± 0.08 | Hurt |
| Set αmax above 0 when sampling adjacent latent codes (ablation) | Local unlearning baseline · Out-of-Domain (CelebAHQ), multi-image test | IDothers: 0.2754 ± 0.0791 -> 0.2123 ± 0.0762 (αmax=10); 0.2094 ± 0.0692 (αmax=15); 0.2105 ± 0.0924 (αmax=20); 0.2111 ± 0.0653 (αma | Helped |
| Add global preservation loss (ablation) | Local unlearning baseline · Out-of-Domain (CelebAHQ) | FIDpre: 9.52 ± 1.53 -> 4.63 ± 0.43 | Helped |
| Add global preservation loss (ablation) | Local unlearning baseline · Out-of-Domain (CelebAHQ) | ΔFIDreal: 4.75 ± 0.89 -> 1.48 ± 0.29 | Helped |
| Add extrapolated target latent, adjacency-aware loss, and global preservation loss (ablation) | Baseline · Random, In-Domain (FFHQ), and Out-of-Domain (CelebAHQ) | ID: 0.19 ± 0.09 -> 0.14 ± 0.06; 0.16 ± 0.07 -> 0.06 ± 0.06; 0.12 ± 0.06 -> 0.03 ± 0.05 | Helped |
Surprising findings
- Adding the adjacency-aware loss alone (without global loss) substantially increases FIDpre (13.63 vs 9.52), i.e., it degrades generation quality, and the global preservation loss is needed to recover it.
Limitations
- Target choice d must be tuned empirically (d=30); d<0 or d=0 fail to fully erase identity, while larger d degrades distribution.
- Only evaluated on EG3D trained on FFHQ with a single inversion network (GOAE).
- Identity removal evaluated only with a single face recognition network (CurricularFace).
Experimental setup
Datasets: FFHQ CelebAHQ
Benchmarks: FFHQ CelebAHQ
Metrics: ID similarity (CurricularFace) IDothers FIDpre ΔFIDreal
Baselines: Baseline (average latent target with local unlearning loss)
Builds on: EG3D GOAE StyleGAN2 ESD
Closest papers by meaning
- Machine Unlearning for Image-to-Image Generative Models (ICLR 2024 · similarity 0.91)
- Controllable Unlearning for Image-to-Image Generative Models via $\epsilon$-Constrained Optimization (ICLR 2025 · similarity 0.89)
- Boosting Alignment for Post-Unlearning Text-to-Image Generative Models (NeurIPS 2024 · similarity 0.87)
- FaceA-Net: Facial Attribute-Driven ID Preserving Image Generation Network (AAAI 2025 · similarity 0.87)
- Benchmarking Vision Language Model Unlearning via Fictitious Facial Identity Dataset (ICLR 2025 · similarity 0.87)
- Feature Unlearning for Pre-trained GANs and VAEs (AAAI 2024 · similarity 0.87)
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.