Research map / Label-efficient and robust learning / Recognition and long-tail
Partitioning for Intrinsic Model Inversion Resistance in Collaborative Inference
Rongke Liu, Youwen Zhu, Lei Zhou, Zhang Xianglong, Dong Wang
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 |
|---|---|---|---|
| partition at decision-level block 49+ instead of earlier blocks | IR-152 · CIFAR-10 MIA | reconstruction MSE: below 0.02 up to block 48, then sharp rise | Improved |
| partition at VGG19 Block 39 instead of Block 30 | VGG19 · CIFAR-10 MIA | reconstruction MSE: 0.066493 -> 0.136896 | Improved |
| increase depth of ViT partition point | ViT (6-depth) · CIFAR-10 MIA | reconstruction MSE | No clear effect |
| train with label smoothing (ablation) | IR-152 / VGG19 / ViT · CIFAR-10 | test accuracy: 90.59% -> 90.76% (IR-152) | No clear effect |
| enrich Z with dropout NN module (concat) (ablation) | IR-152 Block 50 · CIFAR-10 MIA, test set | reconstruction MSE: 0.057010 -> 0.051740 | Hurt |
| enrich Z with FFT (residual) (ablation) | IR-152 Block 50 · CIFAR-10 MIA, test set | reconstruction MSE: 0.057010 -> 0.112854 | Improved |
| enrich Z with NN module without dropout (ablation) | IR-152 Block 50 · CIFAR-10 MIA | attack generalization | Hurt |
| enrich Z with augmentation (ablation) | IR-152 · CIFAR-10 MIA | PSNR gain: +0.4 at Block 50 vs about +1.13 at Block 40 | Mixed |
| use attention-enhanced inversion model (ablation) | IR-152 Block 50 · CIFAR-10 MIA, test set | reconstruction MSE: 0.057010 -> 0.050579 (AttentionAsConv+SE+LSK+MSCA) | Hurt |
| use auxiliary data from the same distribution as target (ablation) | VGG19 · KMNIST reconstruction | reconstruction quality | Improved |
| use dimension normalization in R2c-based GPZ location (ablation) | R2c locator · VGG19 | transition location: 26->30 and 39->43 vs 26->30 and 30->39 | Improved |
| use label smoothing for GPZ location (ablation) | R2c locator · IR-152 / VGG19 | transition zone | No clear effect |
Closest papers by meaning
- Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems (CVPR 2025 · similarity 0.87)
- Sparse Model Inversion: Efficient Inversion of Vision Transformers for Data-Free Applications (ICML 2024 · similarity 0.84)
- The Geometric Wall: Manifold Structure Predicts Layerwise Sparse Autoencoder Scaling Laws (NeurIPS 2026 · similarity 0.84)
- Reimagining Mutual Information for Enhanced Defense against Data Leakage in Collaborative Inference (NeurIPS 2024 · similarity 0.83)
- DRAG: Data Reconstruction Attack using Guided Diffusion (ICML 2025 · similarity 0.83)
- Unextractable Protocol Models: Collaborative Training and Inference without Weight Materialization (NeurIPS 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.
