Research map / Adversarial robustness and security / Black-box and transfer attacks
Label-Only Model Inversion Attacks via Knowledge Transfer
Ngoc-Bao Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, Ngai-man Cheung
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Transfer target-model decisions into T-ACGAN rather than training the surrogate with target-labeled public images or standard ACGAN | Surrogate model for CelebA inversion (FaceNet64 target) · CelebA | Attack accuracy: 5.87 $\pm$ 1.65 -> 42.07 $\pm$ 3.46 | Improved |
| Add data augmentation to balance target-labeled public training data (ablation) | Direct I surrogate training · CelebA | Attack accuracy: 5.87 $\pm$ 1.65 -> 9.60 $\pm$ 2.22 | Improved |
| Add an ensemble of surrogate architectures instead of using one surrogate (ablation) | LOKT · CelebA (FaceNet64; CelebA public) | Attack accuracy: 92.80 $\pm$ 2.59 -> 93.93 $\pm$ 2.78 | Improved |
| Use T-ACGAN's generated, target-labeled data to train a separate surrogate rather than use its discriminator/classifier directly (ablation) | LOKT · CelebA (FaceNet64; CelebA public) | Attack accuracy: 81.00 $\pm $ 4.79 -> 92.80 $\pm $ 2.59 | Improved |
| Use T-ACGAN's generated, target-labeled data to train a separate surrogate rather than its discriminator/classifier directly (ablation) | LOKT · Pubfig83 (FaceNet64; Pubfig83 public) | Attack accuracy: 74.80 $\pm $ 5.93 -> 61.60 $\pm $ 3.58 | Hurt |
| Replace BREPMI's black-box search with LOKT's knowledge-transfer surrogate attack | Label-only model inversion · CelebA (FaceNet64; CelebA public) | Attack accuracy: 73.93 $\pm $ 4.98 -> 93.93 $\pm $ 2.78 | Improved |
| Replace BREPMI's black-box search with LOKT's knowledge-transfer surrogate attack | Label-only model inversion · CelebA (FaceNet64; FFHQ public) | Attack accuracy: 43.00 $\pm$ 5.14 -> 62.07 $\pm$ 3.89 | Improved |
| Use LOKT's knowledge-transfer surrogate attack instead of BREPMI | Label-only model inversion · CelebA with BiDO-HSIC target | Attack accuracy: 37.40 $\pm$ 3.66 -> 60.73 $\pm$ 3.07 | Improved |
| Use LOKT's knowledge-transfer surrogate attack instead of BREPMI | Label-only model inversion · CelebA with MID target | Attack accuracy: 39.20 $\pm$ 4.19 -> 60.33 $\pm$ 4.76 | Improved |
| Use LOKT's discriminator/classifier surrogate rather than BREPMI | Label-only model inversion · CelebA (IR152; CelebA public) | Attack accuracy: 71.47 $\pm$ 5.32 -> 72.07 $\pm$ 4.03 | No clear effect |
| Use generated surrogate model S instead of the discriminator/classifier for the CelebA inversion setting (ablation) | LOKT · CelebA (FaceNet64; FFHQ public) | KNN distance: 1516.18 -> 1437.86 | Improved |
| Use the ensemble surrogate rather than BREPMI | Label-only model inversion · CelebA (FaceNet64; CelebA public; query budget) | Queries (millions): 17.98 -> 12.66 | Improved |
Closest papers by meaning
- A Sample-Level Evaluation and Generative Framework for Model Inversion Attacks (AAAI 2025 · similarity 0.90)
- Model Inversion Robustness: Can Transfer Learning Help? (CVPR 2024 · similarity 0.89)
- Be Careful What You Smooth For: Label Smoothing Can Be a Privacy Shield but Also a Catalyst for Model Inversion Attacks (ICLR 2024 · similarity 0.88)
- OSLO: One-Shot Label-Only Membership Inference Attacks (NeurIPS 2024 · similarity 0.86)
- From Head to Tail: Efficient Black-box Model Inversion Attack via Long-tailed Learning (CVPR 2025 · similarity 0.86)
- Towards Eliminating Hard Label Constraints in Gradient Inversion Attacks (ICLR 2024 · similarity 0.86)
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.
