Research map / Federated learning and privacy / Communication-efficient FL
A High-Efficiency Federated Learning Method Using Complementary Pruning for D2D Communication (Student Abstract)
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 FedCPD, a federated learning method that combines complementary L1 pruning masks between server and clients with chain (device-to-device) client updates to cut communication volume.
Problem
Frequent parameter transmission between clients and the server in federated learning causes large communication overhead due to parameter redundancy, and pruning can wrongly remove important weights.
Approach
Apply complementary pruning on server and clients (the pruned positions are filled back via an inverse-mask Hadamard product and aggregation with the sparse global model), and reduce server communication by chaining updates sequentially across clients over D2D links.
Key results
- On CIFAR-10, FedCPD reaches 91% accuracy at round 90 with 5.3E+03 Mb communication overhead and 4.9 TFLOPs in Table 1.
- On MNIST, FedCPD reaches 98% accuracy at round 90, versus 93% at round 30 with 0.8E+06 Mb overhead.
- The paper claims FedCPD consistently outperforms baselines in test accuracy and loss across datasets (Figure 2), but gives no quantitative baseline comparison in text.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedCPD with complementary pruning and D2D chain updates instead of the compared federated learning algorithms | Federated learning · MNIST, Fashion MNIST, CIFAR-10, and CIFAR-100 | test accuracy | Helped |
| Use FedCPD with complementary pruning and D2D chain updates instead of the compared federated learning algorithms | Federated learning · MNIST, Fashion MNIST, CIFAR-10, and CIFAR-100 | communication volume | Helped |
| Use FedCPD with complementary pruning and D2D chain updates instead of the compared federated learning algorithms | Federated learning · MNIST, Fashion MNIST, CIFAR-10, and CIFAR-100 | FLOPs | Helped |
Limitations
- Experiments are only summarized; no quantitative comparison table against baselines is given in the text.
- Communication overhead values in Table 1 are non-monotonic and unexplained (e.g., MNIST 0.3E+05 at round 10 vs 0.8E+06 at round 30).
Experimental setup
Datasets: MNIST Fashion-MNIST CIFAR-10 CIFAR-100
Benchmarks: MNIST Fashion-MNIST CIFAR-10 CIFAR-100
Metrics: test accuracy average loss communication overhead FLOPs
Baselines: FedAvg FedADP CS HL-SGD ZeroFL
Builds on: FedAvg L1 pruning
Closest papers by meaning
- Device-Wise Federated Network Pruning (CVPR 2024 · similarity 0.87)
- FedRTS: Federated Robust Pruning via Combinatorial Thompson Sampling (NeurIPS 2025 · similarity 0.85)
- SpaFL: Communication-Efficient Federated Learning With Sparse Models And Low Computational Overhead (NeurIPS 2024 · similarity 0.85)
- Communication-Efficient Federated Learning with Accelerated Client Gradient (CVPR 2024 · similarity 0.85)
- Debiasing Federated Learning with Correlated Client Participation (ICLR 2025 · similarity 0.85)
- FedMef: Towards Memory-efficient Federated Dynamic Pruning (CVPR 2024 · similarity 0.85)
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.