Code for work A Robust Training Method for Federated Learning with Partial Participation

Experimental Setup for the Paper

This repository includes all experiments and implementations required for the PPBC_ICML study.

---

✅ Architectures and Datasets

• Model: ResNet-18
• Dataset: CIFAR-10
• Client Configurations: Experiments conducted with 10
• Additional Model Support: FasterVIT with the Food101 dataset

---

✅ Implemented Client Selection Strategies

Method	Command	Description
Top-k via loss	method: 'loss'; k: k	Selects top-k clients with the highest loss values
Top-k via gradient norm	method: 'gradient_norm'; k: k	Selects top-k clients with the largest gradient norms
Top-k via BANT method	method: 'bant'; k: k	Uses BANT trust scores to select k clients with the highest trust
Top-k via random sampling	method: 'random'; k: k	Randomly selects k clients on each training round

---

✅ Additional Strategies

Method	Description
Hybrid top-k via method1 + top-t via method2	Selects k clients using method1 and t additional clients using method2
Top-k via gradient angle similarity	Selects k clients whose gradients have the smallest deviation from the mean gradient direction

---

✅ Command to Launch Experiments

    python utils/cifar_download.py
    python train.py

Citation

@misc{partialparticipation,
  title={A Robust Training Method for Federated Learning with Partial Participation},
  author={Dmitry Bylinkin and Daniil Medyakov and Vladimir Aletov and Nail Bashirov and Aleksandr Beznosikov},
  year={2025},
  howpublished = "[Online]. Available: \url{https://scholar.google.com/citations?view_op=view_citation&hl=ru&user=1hPG3aIAAAAJ&citation_for_view=1hPG3aIAAAAJ:KlAtU1dfN6UC}"
}