Optimizing Communication Efficiency and Ensuring Privacy in Federated Learning via Random Projection

Please enable javascript in your browser.

Kazemi, Mohammad Reza | 2024

0 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 58357 (05)
University: Sharif University of Technology
Department: Electrical Engineering
Advisor(s): Yasaee, Mohammad Hossein
Abstract:
Federated learning (FL) systems, recognized for their decentralized machine learning capabilities across multiple nodes without aggregating data, often grapple with high communication overhead and stringent privacy requirements. This thesis proposes an innovative approach that utilizes random projection and compressed sensing techniques to project gradients in weight spaces into a lower-dimensional subspace, effectively reducing communication overhead to O(logd) bits. This method distinguishes itself by addressing both communication efficiency and privacy simultaneously, a notable advancement over traditional methods that typically focus on one at the expense of the other. By implementing differential privacy within this framework, our approach not only minimizes the size of transmitted gradients but also ensures robust privacy protection. Comprehensive experimental results affirm that our methodology substantially lowers communication costs while adhering to privacy constraints, demonstrating its potential to facilitate more scalable and privacy-preserving federated learning implementations
Keywords:
Federated Learning ; Compressive Sensing ; Privacy ; Random Projection ; Communication Overhead ; Privacy Preserving

No TOC