Abstract:
Modern federated and distributed machine learning systems are vulnerable to different types of failures. One natural and powerful class of attacks is the Byzantine attack, where some nodes may behave arbitrarily and even collude.
These nodes may know the data, the algorithm, and all its parameters, including those used by the center, and can use this information to design worst case attacks.
In this talk, we analyze secure (Byzantine resilient) algorithms for solving three federated low rank matrix learning problems that share common features: low rank matrix completion, multi task representation learning, and robust principal component analysis (RPCA). These problems appear in many modern
machine learning and medical imaging applications, such as recommender systems, federated sketching, accelerated dynamic MRI, and Fourier ptychography.
A newer set of applications where the RPCA model is becoming popular is large language model (LLM) compression. Although powerful, LLMs are expensive to run on edge or consumer devices, which makes strong compression necessary. This naturally leads to an RPCA view of the problem.
The algorithms discussed in this talk are secure and communication efficient, making them practical for large scale federated deployment.
Bio
Ankit Pratap Singh is a PhD candidate in the Department of Electrical and Computer Engineering at Iowa State University. His research focuses on federated and secure machine learning and has led to publications at ICML 2024 and 2025, among others. He was awarded the Research Excellence Award in 2025 by Iowa State University. He received a Master’s degree in Statistics and Computing from Banaras Hindu University (BHU), India, in 2020, where he was awarded the Gold Medal for securing first position in the MSc Statistics and Computing program. He has worked as a Machine Learning Researcher at the International Finance Corporation (World Bank Group), where he developed agentic AI
systems. He has also worked as an AI Researcher at Thoughtworks, focusing on LLM driven systems, and synthetic data generation pipelines