PhD Thesis Proposal - Tingting Tang

Title: Optimizing Privacy-Preserving Machine Learning for Improved Privacy, Utility, and Efficiency Tradeoffs    
 
Location: EEB 349  
 
Date and Time: February 3, 2025, 12.30 PM-1.30 PM    
 
Zoom Link: https://usc.zoom.us/j/7995244109?pwd=OUp6RWhUZlFGclgyN3hkREh0Z21ldz09    
 
Committee: Murali Annavaram (Chair), Salman Avestimehr, Bhaskar Krishnamachari, Harsha Madhyastha, Sai Praneeth Karimireddy    
 
Abstract:  Privacy-preserving machine learning (PPML) is essential for protecting sensitive data in machine-learning applications, requiring a careful balance between privacy, utility, and efficiency. However, the trade-offs and interdependencies among these dimensions present significant design challenges. This thesis proposal explores and optimizes their interplay through low-rank decomposition, focusing on two key PPML technologies: Differential Privacy (DP) and Secure Multiparty Computation (MPC). In the context of DP-based graph neural networks (GNNs), I propose a novel training framework leveraging low-rank singular value perturbation to protect sensitive graph edges while preserving the primary graph structure. This approach achieves a significantly improved privacy-utility trade-off and demonstrates resilience to edge inference attacks. For MPC-based secure model inference, I propose leveraging low-rank decomposition for the linear layers of ML models, reducing the number of MPC multiplications required during offline and online phases. Techniques such as truncation skipping and linear layer concatenation further reduce computational and communication overheads, enhancing overall efficiency in MPC ML workflows without compromising the robust security guarantees provided by MPC. By addressing the interactions between privacy, utility, and efficiency, my proposal lays the foundation for more practical and effective deployment of privacy-preserving machine learning solutions in real-world applications.

Location: Hughes Aircraft Electrical Engineering Center (EEB) - 349

Audiences: Everyone Is Invited

Contact: Tingting Tang