Figure 1: The Best Paper (Runner-Up) Award badge

Recently, the paper “HtFLlib: A Comprehensive Heterogeneous Federated Learning Library and Benchmark”—authored by Jianqing Zhang, a PhD student at the Institute of Intelligent Software and Systems of the School of Computer Science at Shanghai Jiao Tong University, with Professor Jian Cao as the corresponding author—received the Best Paper (Runner-Up) Award in the “Datasets and Benchmarking” track at KDD 2025 (with only three papers selected in this category). Standing out among numerous submissions, the paper was recognized by reviewers for filling a critical gap in standardized evaluation frameworks within the field, highlighting the research team’s strong academic influence in federated learning.

Introduction

The paper highlights the open-source project HtFLlib (project link: https://github.com/TsingZ0/HtFLlib), initiated and led by Jianqing Zhang. HtFLlib is a federated learning framework that enables collaborative training among clients using heterogeneous models. In real-world applications, due to differences in computing power, storage capacity, or specific model-size requirements of different tasks, clients often need to adopt models with varying architectures. However, existing federated learning frameworks generally lack systematic support for such heterogeneous scenarios. HtFLlib fills this gap by introducing a lightweight information exchange mechanism that allows models with diverse architectures to collaborate efficiently within a unified platform.

Currently, HtFLlib supports 40 heterogeneous models, 3 data modalities (images, text, and sensor data), and 10 mainstream heterogeneous federated learning algorithms. It is also highly extensible—users can integrate new algorithms quickly by modifying only two configuration files.

Figure 2: Overview of the HtFLlib project

Under the supervision of Professor Jian Cao, Jianqing Zhang has been dedicated to research in personalized federated learning. He led the development of another open-source project, PFLlib (https://www.pfllib.com/), which focuses on providing a unified evaluation framework for federated learning in data-heterogeneous scenarios. PFLlib includes 39 federated learning algorithms, 3 types of data heterogeneity settings, and 24 datasets. It supports rapid simulation of training scenarios with up to 500 devices and offers tools for assessing privacy-preserving capabilities. To date, PFLlib has received over 1,800 stars and 300 forks on GitHub. Together, PFLlib and HtFLlib have become important foundational platforms in the field and have been adopted by various research and industry institutions worldwide. Thanks to the high usability and extensibility of these two platforms, Jianqing Zhang has published nine CCF-A papers as the first author, achieving significant academic impact.

The research group is currently advancing the deployment of HtFLlib in edge computing environments, aiming to enable “simultaneous data collection and model training.” The system has successfully executed heterogeneous collaborative model training across 48 types of heterogeneous devices—including various microcontrollers and smartphones—further extending the applicability of federated learning in resource-constrained settings.