Research map / Federated learning and privacy / Privacy attacks and secure learning
Revealing and Protecting Labels in Distributed Training
Trung Dang, Om Thakkar, Swaroop Ramaswamy, Rajiv Mathews, Sang Chin, Françoise Beaufays
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 |
|---|---|---|---|
| Use RLG instead of Yin et al.'s method | EfficientNetB0 label-set inference · ImageNet validation, N=10/50/100 | F1: .020 -> 1; .093 -> .999; .174 -> .999 | Improved |
| Use RLG instead of Yin et al.'s method | ResNet50 label-set inference · ImageNet validation, N=10/50 | F1: .998 -> .998; .987 -> .987 | No clear effect |
| Use RLG instead of Yin et al.'s method | ResNet50 label-set inference · ImageNet validation, N=100 | F1: .979 -> .972 | Hurt |
| Use RLG on an untrained model's projection-layer update | ASR bag-of-words inference · LibriSpeech test | EM: .018 -> .988 | Improved |
| Train the ASR model for 4k rather than use the untrained model (ablation) | RLG bag-of-words inference · LibriSpeech test | EM: .988 -> .955 | Hurt |
| Train the ASR model for 10k rather than use the untrained model (ablation) | RLG bag-of-words inference · LibriSpeech test | EM: .988 -> .933 | Hurt |
| Use a single-step mini-batch update instead of aggregating multiple steps (ablation) | RLG bag-of-words inference · LibriSpeech test, N=4 or 8, K=1 | EM: .535 -> 1; .160 -> 1 | Improved |
| Aggregate single-sample updates over 4 or 8 steps (ablation) | RLG bag-of-words inference · LibriSpeech test | EM: 1 -> .535; 1 -> .160 | Hurt |
| Aggregate 2-sample updates over 2 steps (ablation) | RLG bag-of-words inference · LibriSpeech test, N=2, K=2 | EM: 1 -> .330 | Hurt |
| Apply Sign-SGD to the projection-layer update (ablation) | RLG · EfficientNetB0 and ASR | EM: 1 -> 0; .988 -> .104 | Hurt |
| Apply GradDrop at 50% (ablation) | RLG · EfficientNetB0 and ASR | EM: 1 -> .17; .988 -> .881 | Mixed |
| Apply GradDrop at 90% (ablation) | RLG · EfficientNetB0 and ASR | EM: 1 -> 0; .988 -> .866 | Mixed |
Closest papers by meaning
- Recovering Labels from Local Updates in Federated Learning (ICML 2024 · similarity 0.90)
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.86)
- Towards Eliminating Hard Label Constraints in Gradient Inversion Attacks (ICLR 2024 · similarity 0.86)
- DAGER: Exact Gradient Inversion for Large Language Models (NeurIPS 2024 · similarity 0.86)
- Little Is Enough: Boosting Privacy by Sharing Only Hard Labels in Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.86)
- Retraining with Predicted Hard Labels Provably Increases Model Accuracy (ICML 2025 · 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.
