Organizer：何俊材

Abstract: Federated learning (FL) enables fine-tuning of large language models (LLMs) on decentralized data, but faces two critical challenges: high communication costs and privacy risks. This talk presents two novel methods to address these issues. First, we introduce CG-FedLLM, a communication-efficient FL framework that compresses gradients via a client-side encoder and server-side decoder. Its two-phase training includes gradient-aware pre-training and autoencoder-assisted fine-tuning, significantly reducing bandwidth. Second, we propose a two-stage randomness method named DR-Encoder for end-to-end privacy protection. It combines a Gaussian-prior gradient autoencoder with noise-injected fine-tuning, rigorously analyzed under Gaussian and Rényi differential privacy. Evaluations across multiple LLMs and benchmarks demonstrate improved efficiency, maintained accuracy, and strong theoretical privacy guarantees.

Bio: Dr. Huiwen Wu received her Ph.D. from the University of California, Irvine in 2019. Following her doctoral studies, she served as a Senior Algorithm Engineer at Ant Group and Zhejiang Lab, where she led the design and development of privacy-preserving machine learning algorithms, contributing to cutting-edge advancements in secure and scalable AI systems. In 2026, she joined Beihang University as an associate researcher,  further advancing her research at the intersection of machine learning, data privacy, and optimization. Her research focuses on privacy-preserving machine learning, randomized optimization methods, and random matrix theory, with a strong emphasis on both theoretical foundations and real-world applications. Her work has been published in several top-tier conferences, including The Web Conference (WWW), AAAI, and IJCAI.