Events Calendar

PhD Dissertation Defense - Tingting Tang

Title: Optimizing Privacy-Preserving Machine Learning for Improved Privacy, Utility, and Efficiency Tradeoffs    
 
Date and Time: June 25, 2025, 2 PM-4 PM    
 
Location: EEB 539    
 
Zoom link:https://usc.zoom.us/j/7995244109?pwd=OUp6RWhUZlFGclgyN3hkREh0Z21ldz09
 
Committee: Murali Annavaram (Chair), Bhaskar Krishnamachari, Sai Praneeth Karimireddy, Mengyuan Li    
 
Abstract: Machine learning (ML) has become integral to modern data-driven systems, supporting applications from image recognition and recommendation to drug discovery and language modeling. However, the growing reliance on sensitive data during both training and inference raises critical privacy concerns. Training datasets often contain personal or proprietary information, such as social networks or medical records, while inference pipelines, especially those involving retrieval-augmented generation (RAG), may expose confidential retrieved documents or context. These risks are further exacerbated when ML models are deployed in untrusted cloud environments, making privacy-preserving machine learning (PPML) a pressing research challenge.  
 
This dissertation investigates practical approaches to PPML, focusing on improving the tradeoff between privacy, utility, and efficiency in frameworks based on differential privacy (DP) and secure multiparty computation (MPC). For DP-based PPML, we first introduce a training algorithm for graph neural networks 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 inference-time privacy, we turn to RAG systems and propose a differentially private algorithm that extracts the most frequent keywords in the ensemble of responses using private aggregation. These keywords are then used to construct prompts to produce final responses, reducing the risk of information leakage while maintaining output quality. For MPC-based secure model inference, we present a low-rank decomposition framework that reduces the number of secure multiplications in linear layers. Two further optimizations, truncation skipping and layer concatenation, reduce overhead and improve efficiency across both 3-PC and n-PC protocols.
 
Together, these contributions advance the practical deployment of PPML by offering techniques that uphold formal privacy guarantees while maintaining strong model performance and system efficiency. Through a combination of low-rank approximation, semantic compression, and protocol-aware system optimizations, this dissertation offers a practical path forward for developing privacy-preserving ML systems for real-world deployment.

Location: Hughes Aircraft Electrical Engineering Center (EEB) - 539

Audiences: Everyone Is Invited

Contact: Tingting Tang

This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor.

Add to Google Calendar