Research map / Adversarial robustness and security / Black-box and transfer attacks
Black-Box Privacy Attacks on Shared Representations in Multitask Learning
John Abascal, Nicolás Berrios, Alina Oprea, Jonathan Ullman, Adam Smith, Matthew Jagielski
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 |
|---|---|---|---|
| give adversary the actual training samples (strong) instead of fresh samples (weak) | inner product attack · CelebA personalization, 50th percentile | balanced accuracy: 52.4% -> 80% | Improved |
| strong instead of weak adversary | inner product attack · Stack Overflow personalization, 50th percentile | balanced accuracy: 58.2% -> 98.7% | Improved |
| strong instead of weak adversary | variance attack · CelebA personalization, FPR 1% | TPR: 2.9% -> 61.2% | Improved |
| strong instead of weak adversary | inner product attack · Stack Overflow personalization, FPR 1% | TPR: 8.2% -> 98.5% | Improved |
| define tasks as labelled topics (multiple learning problems) instead of users | weak adversary inner product attack · Stack Overflow, 50th percentile | balanced accuracy: 58.2% -> 85.9% | Improved |
| define tasks as users (personalization) rather than labelled topics | strong vs weak gap · Stack Overflow | balanced accuracy: 98.7% -> 87% | Mixed |
| raise percentile threshold from 50th to 90th | inner product attack strong adversary · CelebA personalization | balanced accuracy: 80% -> 59.5% | Hurt |
| increase total samples (samples per task or number of tasks) (ablation) | variance attack strong adversary · synthetic data | AUC | Improved |
| increase embedding dimension (ablation) | variance attack · synthetic data | AUC | No clear effect |
| larger generalization gap (train loss minus zero-shot loss) | strong adversary inner product attack · Stack Overflow | AUC | Improved |
| use gradient clipping and normalization as defense | MTL training · task-inference attacks | attack success | No clear effect |
Closest papers by meaning
- MTL-UE: Learning to Learn Nothing for Multi-Task Learning (ICML 2025 · similarity 0.89)
- Identifying and Mitigating Spurious Correlation in Multi-Task Learning (CVPR 2025 · similarity 0.85)
- Adversarially Robust Multi-task Representation Learning (NeurIPS 2024 · similarity 0.85)
- Distribution Matching for Multi-Task Learning of Classification Tasks: A Large-Scale Study on Faces & Beyond (AAAI 2024 · similarity 0.85)
- Privacy Backdoors: Enhancing Membership Inference through Poisoning Pre-trained Models (NeurIPS 2024 · similarity 0.84)
- Learning Task-Agnostic Representations through Multi-Teacher Distillation (NeurIPS 2025 · similarity 0.84)
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.
