Research map / Federated learning and privacy / Communication-efficient FL
Understanding Server-Assisted Federated Learning in the Presence of Incomplete Client Participation
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.
Shows theoretically that conventional FL is not PAC-learnable under incomplete client participation, that server-assisted FL (SA-FL) restores PAC-learnability, and proposes SAFARI, a provably convergent SA-FL algorithm evaluated on MNIST and CIFAR-10.
Problem
Federated learning when some clients never participate (incomplete client participation), where the participating data distribution deviates from the population distribution; theoretical understanding of server-assisted FL (SA-FL) and its training convergence is lacking.
Approach
Model SA-FL with a server-side auxiliary dataset drawn from the population distribution, showing a worst-case impossibility for plain FL and a PAC bound for SA-FL via a mixture distribution with the same support; the SAFARI algorithm randomly alternates FedAvg-style client rounds with SGD server rounds, controlled by a probability q.
Key results
- FedAvg test accuracy drops up to 27.44% on MNIST and 25% on CIFAR-10 when s=4 clients are excluded under non-i.i.d. data (p=1).
- SAFARI with q=0.8 improves MNIST accuracy by up to 31.07% over FedAvg at p=1, s=4 with 1000 server samples.
- SAFARI achieves O(1/sqrt(mKR)) convergence for non-convex and tilde-O(1/R) for strongly convex objectives, matching classic FL rates.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| exclude more clients from participation (s=0 to s=4) | FedAvg · MNIST, p=1 | test accuracy (%): 84.49 -> 57.05 | Hurt |
| exclude more clients from participation (s=0 to s=4) | FedAvg · CIFAR-10, p=1 | test accuracy (%): 75.7 -> 50.7 | Hurt |
| add server auxiliary data | FedAvg / SAFARI · MNIST and CIFAR-10, homogeneous cases (p=10 to p=5) | test accuracy | No clear effect |
| add server auxiliary data | FedAvg / SAFARI · MNIST and CIFAR-10 with incomplete participation and heterogeneous data | test accuracy improvement (%): +31.07 | Helped |
| increase server auxiliary data size from 50 to 1000 (ablation) | SAFARI (q=0.8) · MNIST, s=4, p=1 | test accuracy improvement (%): 16.65 -> 31.07 | Helped |
Surprising findings
- Conventional FL is not PAC-learnable under incomplete client participation in the worst case, with a lower bound independent of per-client sample size.
- Under IID data, excluding clients barely changes FedAvg accuracy, while non-IID data causes large drops.
Limitations
- Convergence guarantees rely on conditions on q and the server-client gradient ratio; the rate only holds under specific round ratios.
- Experiments are small-scale (logistic regression and a small CNN with 10 clients).
- Theory assumes the auxiliary dataset is i.i.d. from the population distribution, with relaxation only discussed in remarks.
Experimental setup
Datasets: MNIST CIFAR-10
Benchmarks: MNIST CIFAR-10
Metrics: test accuracy
Baselines: FedAvg
Builds on: FedAvg domain adaptation theory Bernstein class condition
Closest papers by meaning
- Exact and Linear Convergence for Federated Learning under Arbitrary Client Participation is Attainable (NeurIPS 2025 · similarity 0.89)
- Federated Learning under Periodic Client Participation and Heterogeneous Data: A New Communication-Efficient Algorithm and Analysis (NeurIPS 2024 · similarity 0.89)
- Debiasing Federated Learning with Correlated Client Participation (ICLR 2025 · similarity 0.88)
- A Lightweight Method for Tackling Unknown Participation Statistics in Federated Averaging (ICLR 2024 · similarity 0.87)
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.87)
- Low Precision Local Training is Enough for Federated Learning (NeurIPS 2024 · similarity 0.87)
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.