Transparency and Explainability of Federated Learning

🎯 Overview

This research project focuses on Transparency and Explainability of Federated Learning, addressing critical challenges in making federated learning systems more interpretable, cost-aware, and adaptable to domain shifts. The work is conducted by research teams at Tan Tao University (Vietnam) and National Institute of Information and Communications Technology (Japan).

🌟 Research Highlights

Our research tackles fundamental challenges in explainable and transparent federated learning:

Cost-Performance Analysis: Framework for evaluating trade-offs between training costs and model performance
Privacy-Preserving Contribution Measurement: Quantifying client contributions without accessing raw datasets
Intelligent Client Selection: Algorithms that balance performance gains with training expenses
Domain Adaptation with Server Feedback: Lightweight feedback mechanisms (Grad-CAM, SHAP) for better generalization
Modality-Specific Data Generation: TVAE for tabular data, diffusion models for images, LLMs for text

🚀 Key Features

1. Explainable Cost-Performance Framework

Comprehensive transparency in federated learning economics:

Four-Component Cost Model: Data usage, market context, model improvement, resource consumption
Privacy-Preserving Measurement: Analyzes model behavior via Wasserstein distance, not raw data
Interpretable Dashboards: Visual analytics for cost distribution and performance ratios
Real-Time Monitoring: Tracks contributions and expenses throughout training

2. Intelligent Client Selection

Strategic optimization for Quality of Training:

Two-Phase Approach: Clustering (K-Means + K-Medoids) followed by greedy selection
Alternating Criteria: Balances performance and cost across training rounds
Fairness Mechanism: Penalty function ensures all clients get opportunities
Proven Results: 8-32% cost reduction, 7-26% time savings while maintaining accuracy

3. Domain Adaptation with Server Feedback

Novel approach for handling distributional shifts:

Server-Side Analysis: Detects domain shifts and identifies misclassified samples
Lightweight Feedback: Grad-CAM (images) and SHAP (tabular) signals without raw data access
Client-Side Adaptation: Targeted training guided by interpretable feedback
Significant Gains: Up to 10% accuracy improvement under domain shift

Privacy-preserving synthetic data augmentation:

Tabular: TVAE with feedback-guided feature injection
Image: Diffusion models conditioned on Grad-CAM regions
Text: LLM-based generation with validation criteria
Enhanced Performance: Outperforms baselines across all modalities

📚 Publications

Published Papers

1. Client Selection of Federated Learning by Cost-Performance Analysis

Authors: Tien-Dung Cao, Hoang-Duc Le
Venue: 20th International Conference on Intelligent Systems and Knowledge Engineering (ISKE 2025)
Pages: 209-216
Publisher: IEEE

Keywords: Federated Learning, Client Selection, Cost Model, Quality of Training, Explainable AI

2. Domain Adaptation of Federated Learning by Data Generation and Server Feedback

Authors: Phuong-Anh Vu, Kim-Tinh Phan, Cao-Dien Nguyen, Tien-Dung Cao, Le Trieu Phong, Ngoc-Thai Nguyen
Venue: Multi-disciplinary International Workshop on Artificial Intelligence (MIWAI 2025)
Series: Lecture Notes in Artificial Intelligence (LNAI), vol 16354
Pages: 272-283
DOI: 10.1007/978-981-95-4960-3_22
Publisher: Springer Nature Singapore

Keywords: Federated Learning, Server Feedback, Data Generation, Knowledge Distillation, Domain Adaptation

Research Impact

Publications: 2 conference papers (2025)
Venues: IEEE ISKE, Springer MIWAI (LNAI)

👥 Research Team

Principal Investigator

Tien-Dung Cao, PhD
Dean, School of Information Technology
Tan Tao University, Vietnam
📧 dung.cao@ttu.edu.vn

Collaborators

Le Trieu Phong, PhD
National Institute of Information and Communications Technology (NICT), Tokyo, Japan
📧 phong@nict.go.jp

Ngoc-Thai Nguyen
Viettel Solutions, Ho Chi Minh City, Vietnam

Hoang-Duc Le
Faculty of Mathematics & Computer Science, University of Science, VNU-HCMC, Vietnam
📧 lhduc@hcmus.edu.vn

Students

Phuong-Anh Vu, Kim-Tinh Phan, Cao-Dien Nguyen
School of Information Technology, Tan Tao University, Vietnam

🙏 Acknowledgments

This research is supported by:

Tan Tao University Foundation for Science and Technology Development
Grant No. TTU.RS.25.102.003
JST CREST, Japan
Grant No. JPMJCR21M1 (for work of L. T. Phong)

📜 License

This project is licensed under the MIT License.

📞 Contact & Resources

Contact

📧 dung.cao@ttu.edu.vn
🏛️ Tan Tao University, Long An, Vietnam

Resources

📖 Flower Framework
📖 Grad-CAM
📖 SHAP

🔖 Citation

@inproceedings{cao2025client,
  title={Client Selection of Federated Learning by Cost-Performance Analysis},
  author={Cao, Tien-Dung and Le, Hoang-Duc},
  booktitle={2025 20th International Conference on Intelligent Systems and Knowledge Engineering (ISKE)},
  pages={209--216},
  year={2025},
  organization={IEEE}
}

@inproceedings{vu2025domain,
  title={Domain Adaptation of Federated Learning by Data Generation and Server Feedback},
  author={Vu, Phuong-Anh and Phan, Kim-Tinh and Nguyen, Cao-Dien and Cao, Tien-Dung and Phong, Le Trieu and Nguyen, Ngoc-Thai},
  booktitle={Multi-disciplinary International Workshop on Artificial Intelligence},
  pages={272--283},
  year={2025},
  organization={Springer}
}

Transparency and Explainability of Federated Learning

Making Federated Learning More Interpretable, Cost-Aware, and Adaptable

Research by Tan Tao University and Collaborators