Research map / Federated learning and privacy / Privacy attacks and secure learning
MaSS: Multi-attribute Selective Suppression for Utility-preserving Data Transformation from an Information-theoretic Perspective
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 MaSS, an information-theoretic learnable data transformation that suppresses multiple sensitive attributes while preserving annotated and unannotated useful attributes via adversarial CE constraints and an InfoNCE mutual-information term.
Problem
Utility-preserving privacy protection: transforming datasets so that multiple sensitive attributes cannot be inferred while useful annotated and unannotated attributes remain predictable, with theoretical grounding.
Approach
Formulates attribute suppression/preservation as a constrained mutual-information problem, relaxes constraints into differentiable penalty losses using adversarial cross-entropy estimators, and preserves unannotated features by maximizing an InfoNCE lower bound on I(X';F) instead of l2 reconstruction.
Key results
- On Motion Sense (suppress gender and ID, preserve activity as unannotated), MaSS reaches NAG 0.0000 on gender and 0.8977 on activity, the highest among all methods.
- On AudioMNIST (suppress gender, accent, age, ID; digit unannotated), MaSS achieves digit NAG 0.9675 with ID NAG 0.0029, versus GAP digit NAG 0.9485 and ID 0.0281.
- On Adience (suppress gender), MaSS gets gender NAG 0.0000 with age NAG 0.7418 and ID NAG 0.7662, while MSDA reaches higher age and ID NAG but leaks gender (0.3114).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MaSS instead of MSDA for preserving unannotated activity | Motion Sense transformation · Motion Sense (gender and ID suppressed; activity unannotated) | NAG (activity): 0.8418 -> 0.8977 | Helped |
| Replace InfoNCE with ℓ₂ reconstruction loss (ablation) | MaSS · Motion Sense (gender and ID suppressed; activity unannotated) | NAG (activity): 0.8977 -> 0.8156 | Hurt |
| Remove the InfoNCE loss (ablation) | MaSS · Motion Sense (gender and ID suppressed; activity unannotated) | NAG (activity): 0.8977 -> 0.7275 | Hurt |
| Replace InfoNCE with MINE-based negative mutual information (ablation) | MaSS · Motion Sense (gender and ID suppressed; activity unannotated) | NAG (activity): 0.8977 -> 0.6847 | Hurt |
| Replace InfoNCE with MINE-based negative mutual information (ablation) | MaSS · Motion Sense (gender and ID suppressed; activity unannotated) | NAG (gender): 0.0000 -> 0.3294 | Hurt |
| Use MaSS instead of GAP for preserving unannotated digit | AudioMNIST transformation · AudioMNIST (gender, accent, age, and ID suppressed; digit unannotated) | NAG (digit): 0.9485 -> 0.9675 | Helped |
| Replace InfoNCE with ℓ₂ reconstruction loss (ablation) | MaSS · AudioMNIST (gender, accent, age, and ID suppressed; digit unannotated) | NAG (digit): 0.9675 -> 0.9517 | Hurt |
| Remove the InfoNCE loss (ablation) | MaSS · AudioMNIST (gender, accent, age, and ID suppressed; digit unannotated) | NAG (digit): 0.9675 -> 0.1846 | Hurt |
| Replace InfoNCE with MINE-based negative mutual information (ablation) | MaSS · AudioMNIST (gender, accent, age, and ID suppressed; digit unannotated) | NAG (digit): 0.9675 -> 0.5031 | Hurt |
| Replace InfoNCE with MINE-based negative mutual information (ablation) | MaSS · AudioMNIST (gender, accent, age, and ID suppressed; digit unannotated) | NAG (ID): 0.0029 -> 0.0434 | Hurt |
| Use MaSS instead of MSDA to preserve unannotated ID | AudioMNIST transformation · AudioMNIST (gender, accent, and age suppressed; digit annotated; ID unannotated) | NAG (ID): 0.8428 -> 0.8514 | Helped |
| Use MaSS instead of MSDA to preserve unannotated age | Adience transformation · Adience (gender suppressed; age and ID unannotated) | NAG (age): 0.7928 -> 0.7418 | Hurt |
Surprising findings
- The l2 reconstruction-based ablation (MaSS-l2) and MINE-based variant performed worse on several attributes, with MINE leaking gender NAG 0.3294 on Motion Sense.
- PPDAR attains near-zero NAG on activity-preserving tasks (0.6912 on Motion Sense activity), showing that suppression-only methods can destroy unannotated utility.
Limitations
- Relies on the assumption that annotated attributes are deterministic functions of X (degenerate distributions).
- Evaluated on three relatively small benchmarks with limited baselines; no variance reported.
- Lambda fixed to 1 and constraint values chosen manually; penalty relaxation only approximates the constrained problem.
Experimental setup
Datasets: AudioMNIST Motion Sense Adience WLFW
Benchmarks: AudioMNIST Motion Sense Adience
Metrics: normalized accuracy gain (NAG) classification accuracy
Baselines: ALR GAP MSDA BDQ PPDAR SPAct MaSS-NF MaSS-l2 MaSS-MINE
Builds on: ALR PPDAR BDQ GAP MSDA InfoNCE MINE
Closest papers by meaning
- PASS: Private Attributes Protection with Stochastic Data Substitution (ICML 2025 · similarity 0.88)
- Learning Robust and Privacy-Preserving Representations via Information Theory (AAAI 2025 · similarity 0.87)
- DECO-Bench: Unified Benchmark for Decoupled Task-Agnostic Synthetic Data Release (NeurIPS 2024 · similarity 0.86)
- A Certified Unlearning Approach without Access to Source Data (ICML 2025 · similarity 0.86)
- Does Training with Synthetic Data Truly Protect Privacy? (ICLR 2025 · similarity 0.86)
- Disguise without Disruption: Utility-Preserving Face De-identification (AAAI 2024 · similarity 0.85)
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.