Research map / Federated learning and privacy / Communication-efficient FL
LiD-FL: Towards List-Decodable 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 LiD-FL, a list-decodable federated learning framework where the server maintains a list of candidate models so that at least one performs well even when adversaries may exceed half of the clients.
Problem
Federated learning under Byzantine attacks where the fraction of malicious clients may exceed one half, so standard robust aggregators fail to identify honest updates.
Approach
Replace single-model robust aggregation with list decoding: the server keeps a list of q models, with at least one guaranteed to be good, and selects among candidates via a voting/validation step, avoiding dependence on an accurate honest fraction.
Key results
- LiD-FL (ListDec) achieves worst-case accuracy of about 0.67-0.72 across Empire, Gauss, LF, LIE, OMN, SF, and Wst attacks on FEMNIST.
- Baselines CWM, FedAvg, GM and Norm collapse to near-zero accuracy under most attacks (e.g., 0.00-0.02 on Empire, Gauss, OMN, SF, Wst).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ListDec instead of the strongest listed baseline (CWM, FedAvg, or GM) | Federated learning with CNN on Non-IID FEMNIST · Empire | ACC: 0.02 -> 0.69 | Helped |
| Use ListDec instead of the strongest listed baseline (CWM, FedAvg, or GM) | Federated learning with CNN on Non-IID FEMNIST · Gauss | ACC: 0.02 -> 0.68 | Helped |
| Use ListDec instead of FedAvg | Federated learning with CNN on Non-IID FEMNIST · LF | ACC: 0.33 -> 0.67 | Helped |
| Use ListDec instead of Norm | Federated learning with CNN on Non-IID FEMNIST · LIE | ACC: 0.06 -> 0.72 | Helped |
| Use ListDec instead of the strongest listed baseline (CWM, FedAvg, or GM) | Federated learning with CNN on Non-IID FEMNIST · OMN | ACC: 0.02 -> 0.68 | Helped |
| Use ListDec instead of the strongest listed baseline (CWM, FedAvg, GM, or Norm) | Federated learning with CNN on Non-IID FEMNIST · SF | ACC: 0.02 -> 0.67 | Helped |
| Use ListDec instead of the strongest listed baseline (CWM, FedAvg, GM, or Norm) | Federated learning with CNN on Non-IID FEMNIST · Wst | ACC: 0 -> 0.67 | Helped |
Surprising findings
- Standard robust aggregators and preprocessing (NNM, Bucketing) fail when adversaries exceed half of clients, since malicious clients can submit shifted copies of honest updates.
Limitations
- Experiments only on Non-IID FEMNIST; IID setting results were omitted.
- Relies on Definition 4.2 (validation oracle) as an assumption justified mainly by Hoeffding bound and experiments.
- Theory treats p, q, delta, gamma as constants and requires a model list size q >= floor(1/gamma).
Experimental setup
Datasets: FEMNIST
Benchmarks: FEMNIST
Metrics: worst-case accuracy across attacks
Baselines: CWM FedAvg GM Norm ListDec
Builds on: Byzantine-robust aggregation list-decodable learning
Closest papers by meaning
- A Huber Loss Minimization Approach to Byzantine Robust Federated Learning (AAAI 2024 · similarity 0.86)
- Byzantine-Robust Federated Learning: Impact of Client Subsampling and Local Updates (ICML 2024 · similarity 0.85)
- Efficient Federated Learning against Byzantine Attacks and Data Heterogeneity via Aggregating Normalized Gradients (NeurIPS 2025 · similarity 0.84)
- Privacy-Preserving Federated Convex Optimization: Balancing Partial-Participation and Efficiency via Noise Cancellation (ICML 2025 · similarity 0.84)
- Sharp Gaussian approximations for Decentralized Federated Learning (NeurIPS 2025 · similarity 0.84)
- Towards Trustworthy Federated Learning with Untrusted Participants (ICML 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.