Research map / Federated learning and privacy / Privacy attacks and secure learning
Data Valuation and Detections in Federated Learning
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 FedBary, a privacy-preserving federated method that uses a Wasserstein barycenter and interpolating measures to value FL clients and detect noisy data without training a federated model or sharing raw data.
Problem
Fair, efficient, privacy-preserving valuation of clients and individual data points in federated learning before training, especially when many clients are present and validation data may be unavailable.
Approach
Measures client and datum relevance via Wasserstein distance to a federated Wasserstein barycenter, computed iteratively with interpolating measures shared between server and clients; the calibrated gradient of the distance with respect to each datum flags noisy points, avoiding model training and raw data sharing.
Key results
- On CIFAR-10 with 5 clients, FedBary detects 100% of corrupted-feature samples with positive calibrated gradients, matching the noisy-feature detection of Lava and outperforming LOO and KNNShapley.
- FedBary's evaluation time grows linearly with the number of clients: about 2 minutes for N=5 and 1 hour for N=100 with support S=500, versus hours for exact and approximate Shapley methods at N=10.
- Duplicating the full training set up to three times leaves the computed distance essentially unchanged (39.14 to 39.18), showing robustness to duplication.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedBary instead of the reported client-valuation baselines | Client evaluation · Client evaluation, N=5 and N=10 | elapsed time: ExactFed: 31m vs 1m; GTG: 33s vs 1m; MR: 5m vs 1m; DataSV: 25m vs 1m; ExactFed: 3h20m vs 2m; GTG: 7m vs 2m; MR: 40m vs 2 | Helped |
| Increase FedBary support size from S=100 to S=500 (ablation) | FedBary · Client evaluation, N=5, N=10, N=50 and N=100 | elapsed time: 1m / 2m; 2m / 4m; 14m / 30m; 21m / 1h | Hurt |
| Increase FedBary support size (ablation) | FedBary contribution evaluation · Client evaluation | contribution topology | No clear effect |
| Duplicate the whole training dataset (ablation) | FedBary · CIFAR-10, whole dataset duplication | distance: 39.14; 39.23; 39.18 | No clear effect |
| Duplicate a single datum with a large negative gradient value (ablation) | FedBary · CIFAR-10, one datum duplicated | distance: 40.94; 40.96; 40.99 | Hurt |
| Use FedBary for corrupt feature detection | FedBary data detection · CIFAR-10 and various datasets | detection success rate: 100% | Helped |
| Use FedBary rather than Lava for mislabeled-data detection | Federated data detection · CIFAR-10 mislabeled samples | detection accuracy | Mixed |
| Filter noisy samples using FedBary before federated training | FedAVG · Federated learning with 5 clients | model performance | Helped |
| Evaluate noisy-label data using Wasserstein distance | FedBary data detection · CIFAR-10 mislabeled samples | detection accuracy | Hurt |
Surprising findings
- Label-noise detection is relatively poor because Wasserstein distances change only trivially with the portion of label noise, though FedBary still performs similarly to methods that access raw data.
Limitations
- Label-noise detection is weaker since Wasserstein distances change trivially with label noise portion.
- Privacy guarantee is argued informally rather than formally proven.
- Choice of mixture weights lambda for the target distribution is left for future exploration.
Experimental setup
Datasets: CIFAR-10 Gaussian synthetic distributions
Benchmarks: CIFAR-10
Metrics: percentage of contribution detection ratio test accuracy under point removal elapsed time squared barycenter error
Baselines: GTG-Shapley FedShapley MR/OR DataSV exactFed Lava LOO KNNShapley
Builds on: Lava GTG-Shapley federated Wasserstein barycenter (Dvurechensky et al.) FedAvg
Closest papers by meaning
- Fortifying Federated Learning Towards Trustworthiness via Auditable Data Valuation and Verifiable Client Contribution (CVPR 2025 · similarity 0.89)
- A Bayesian Approach for Personalized Federated Learning in Heterogeneous Settings (NeurIPS 2024 · similarity 0.87)
- Federated Learning via Input-Output Collaborative Distillation (AAAI 2024 · similarity 0.87)
- Interaction-Aware Gaussian Weighting for Clustered Federated Learning (ICML 2025 · similarity 0.86)
- Little Is Enough: Boosting Privacy by Sharing Only Hard Labels in Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.86)
- Bridging Model Heterogeneity in Federated Learning via Uncertainty-based Asymmetrical Reciprocity Learning (ICML 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.