Research map / Federated learning and privacy / Communication-efficient FL
Efficient Federated Learning against Heterogeneous and Non-stationary Client Unavailability
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 FedAPM, a federated learning algorithm that handles heterogeneous, non-stationary client unavailability with O(1) extra memory and computation per client via adaptive innovation echoing and implicit gossiping, with a non-convex convergence guarantee and linear speedup.
Problem
Federated learning under intermittent client availability where each client's availability probability is heterogeneous and varies non-stationarily over time, causing biased global updates for FedAvg; prior remedies either assume stationarity or need memory proportional to clients times model dimens
Approach
Clients echo their local innovation scaled by the number of rounds since their last availability (adaptive innovation echoing) so they catch up on missed computation, and the server delays broadcasting the global model so active clients implicitly gossip through aggregation, equalizing weights without memory of stale gradients or knowledge of availability dynamics.
Key results
- FedAPM reaches 86.1% test accuracy on SVHN under stationary availability, versus 82.4% for FedAvg over active clients and 76.1% for FedAvg over all clients.
- Under the non-stationary staircase dynamic on CIFAR-10, FedAPM achieves 66.0% test accuracy, above FedAU (63.9%) and F3AST (63.3%) and close to memory-aided MIFA (65.3%).
- The paper proves FedAPM converges to a stationary point for non-convex objectives with linear speedup in local steps and active clients, with O(1) extra per-client memory and computation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FedAPM instead of FedAvg as client availability becomes more non-stationary | FedAvg on SVHN · SVHN, p=0.1; availability gamma increases from 0.1 to 0.5 | train and test accuracy | Hurt |
| use FedAPM instead of FedAvg with known availability probabilities | Federated learning · SVHN test, stationary | accuracy: 85.6 -> 86.1 | Helped |
| use FedAPM instead of FedAvg with known availability probabilities | Federated learning · CIFAR-10 test, stationary | accuracy: 63.1 -> 66.3 | No clear effect |
| use FedAPM instead of FedAvg with known availability probabilities | Federated learning · CINIC-10 test, stationary | accuracy: 44.6 -> 47.3 | Helped |
| use FedAPM instead of FedAvg with known availability probabilities | Federated learning · SVHN test, staircase | accuracy: 85.2 -> 85.6 | No clear effect |
| use FedAPM instead of FedAvg with known availability probabilities | Federated learning · CIFAR-10 test, staircase | accuracy: 66.1 -> 66.0 | No clear effect |
| use FedAPM instead of FedVARP | Federated learning · CINIC-10 test, staircase | accuracy: 47.2 -> 46.9 | No clear effect |
| use FedAPM instead of FedAvg with known availability probabilities | Federated learning · SVHN test, sine | accuracy: 86.0 -> 85.6 | Hurt |
| use FedAPM instead of FedAvg with known availability probabilities | Federated learning · CIFAR-10 test, sine | accuracy: 67.3 -> 63.5 | Hurt |
| use FedAPM instead of FedAvg with known availability probabilities | Federated learning · CINIC-10 test, sine | accuracy: 47.4 -> 45.8 | Hurt |
| use FedAPM instead of FedAvg with known availability probabilities | Federated learning · SVHN test, interleaved sine | accuracy: 85.2 -> 84.6 | Hurt |
| use FedAPM instead of FedAvg with known availability probabilities | Federated learning · CIFAR-10 test, interleaved sine | accuracy: 66.5 -> 63.3 | Hurt |
Surprising findings
- FedAPM occasionally beats the memory-heavy MIFA and FedVARP despite using no stored gradients.
- FedAPM remains competitive even in the interleaved sine setting where the Assumption on positive availability is violated because probabilities occasionally reach zero.
Limitations
- Analysis assumes independent and strictly positive unavailability probabilities, which the interleaved sine experiment violates.
- Convergence analysis does not cover correlated or zero-valued availability dynamics.
- No incorporation of variance reduction techniques.
Experimental setup
Datasets: SVHN CIFAR-10 CINIC-10
Benchmarks: SVHN CIFAR-10 CINIC-10
Metrics: train accuracy test accuracy
Baselines: FedAvg over active clients FedAvg over all clients FedAU F3AST FedAvg with known p_i^t MIFA FedVARP
Builds on: FedAvg gossip algorithms MIFA FedVARP
Closest papers by meaning
- Exact and Linear Convergence for Federated Learning under Arbitrary Client Participation is Attainable (NeurIPS 2025 · similarity 0.91)
- Federated Learning under Periodic Client Participation and Heterogeneous Data: A New Communication-Efficient Algorithm and Analysis (NeurIPS 2024 · similarity 0.91)
- FedAdaVR: Adaptive Variance Reduction for Robust Federated Learning under Limited Client Participation (NeurIPS 2026 · similarity 0.90)
- Tackling Intertwined Data and Device Heterogeneities in Federated Learning with Unlimited Staleness (AAAI 2025 · similarity 0.89)
- Momentum-Driven Adaptivity: Towards Tuning-Free Asynchronous Federated Learning (ICML 2025 · similarity 0.89)
- Communication-Efficient Federated Learning with Accelerated Client Gradient (CVPR 2024 · similarity 0.89)
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.