Research map / Federated learning and privacy / Communication-efficient FL
GAS: Generative Activation-Aided Asynchronous Split 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 GAS, an asynchronous split federated learning framework with activation and model buffers on the server, where per-label activation Gaussians are maintained and sampled to rebalance server-side updates against straggler-induced bias.
Problem
Split federated learning suffers from stragglers because client activations and client-side models arrive asynchronously, and buffered activations from resource-rich clients bias server-side model updates.
Approach
Maintain a dynamically weighted Gaussian activation distribution per label on the server, built from uploaded activations, and generate label-balanced synthetic activations to concatenate with buffered real activations when updating the server-side model; client-side models are aggregated from a model buffer.
Key results
- On CIFAR-10 with shard=2, GAS reaches 82.78% test accuracy versus 79.57% for CA^2FL and 72.88% for FedAvg.
- On CINIC-10 with shard=2, GAS reaches 68.32% versus 64.29% for CA^2FL.
- On Fashion-MNIST with shard=2, GAS reaches 90.66% versus 88.32% for CA^2FL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GAS instead of CA²FL | asynchronous federated learning · CIFAR-10 shard=2 | test accuracy (%): 79.57 ± 0.98 -> 82.78 ± 0.58 | Helped |
| use GAS instead of CA²FL | asynchronous federated learning · CIFAR-10 α=0.1 | test accuracy (%): 78.56 ± 0.99 -> 81.72 ± 0.50 | Helped |
| use GAS instead of CA²FL | asynchronous federated learning · CINIC-10 | test accuracy (%): 64.29 ± 1.45 -> 68.32 ± 0.17; 68.42 ± 1.11 -> 70.29 ± 0.27; 64.27 ± 0.78 -> 65.94 ± 1.14; 68.77 ± 0.92 -> 69.36 ± 0.65 | Helped |
| use GAS instead of CA²FL | asynchronous federated learning · Fashion-MNIST | test accuracy (%): 88.32 ± 1.19 -> 90.66 ± 0.20; 89.07 ± 0.58 -> 90.58 ± 0.34 | Helped |
| add generative activations to the server-side update (ablation) | GAS · Fashion-MNIST, shard=2 | gradient dissimilarity | Helped |
| add generative activations to the server-side update (ablation) | GAS · Fashion-MNIST, shard=2 | accuracy | Helped |
| increase local iterations from 10 to 20 (ablation) | GAS · CIFAR-10 shard=2 | test accuracy (%): 63.07 ± 0.08 -> 65.58 ± 0.71 | Helped |
| increase local iterations from 20 to 50 (ablation) | GAS · CIFAR-10 shard=2 | test accuracy (%): 65.58 ± 0.71 -> 62.40 ± 1.97 | Hurt |
| increase local iterations from 10 to 20 (ablation) | GAS · CIFAR-10 α=0.1 | test accuracy (%): 60.68 ± 1.13 -> 63.39 ± 2.01 | Helped |
| increase local iterations from 20 to 50 (ablation) | GAS · CIFAR-10 α=0.1 | test accuracy (%): 63.39 ± 2.01 -> 61.96 ± 3.50 | Hurt |
| move the split from AlexNet L4 to shallower L1 (ablation) | GAS with AlexNet · CIFAR-10 shard=3 | test accuracy (%): 72.02 ± 4.39 -> 82.71 ± 1.73 | Helped |
| move the split from VGG16 L3 to L4 (ablation) | GAS with VGG16 · CIFAR-10 shard=3 | test accuracy (%): 76.36 ± 3.69 -> 79.01 ± 2.92 | Helped |
Surprising findings
- FedBuff underperforms synchronous FedAvg on several settings, e.g. CIFAR-10 shard=2 (69.04% vs 72.88%).
Limitations
- Transmits labels and activations, which risks privacy leakage.
- Shallower split layers produce high-dimensional activations that make distribution estimation harder.
- Gaussian activation modeling is an approximation of true activation distributions.
Experimental setup
Datasets: CIFAR-10 CINIC-10 Fashion-MNIST
Benchmarks: CIFAR-10 CINIC-10 Fashion-MNIST
Metrics: test accuracy
Baselines: FedAvg FedBuff CA^2FL MiniBatch-SFL S^2FL
Builds on: Split Federated Learning FedBuff CA^2FL logit adjustment MiniBatch-SFL
Closest papers by meaning
- FSL-SAGE: Accelerating Federated Split Learning via Smashed Activation Gradient Estimation (ICML 2025 · similarity 0.89)
- Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.89)
- Efficient Asynchronous Federated Learning with Prospective Momentum Aggregation and Fine-Grained Correction (AAAI 2024 · similarity 0.88)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.88)
- Towards Straggler-Resilient Split Federated Learning: An Unbalanced Update Approach (NeurIPS 2025 · similarity 0.87)
- Reviving Stale Updates: Data-Free Knowledge Distillation for Asynchronous Federated Learning (NeurIPS 2026 · 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.