Research Assistant at USC

Jan 1, 2024 min read

Overview

As a Research Assistant at USC, I conducted cutting-edge research on privacy-preserving techniques for large language models. My work focused on enabling the benefits of powerful LLMs while protecting sensitive training data.

Research Focus: Privacy-Preserving Knowledge Distillation

The Problem

Large language models like GPT-2 are trained on massive datasets that may contain sensitive information. When deploying these models or sharing them, there’s a risk of inadvertently leaking private data from the training set.

My Approach

I researched privacy-preserving knowledge distillation — a technique that transfers knowledge from a large “teacher” model to a smaller “student” model while incorporating privacy guarantees.

Key components of my research:

Model Alignment

Aligned attention and hidden layers between:

  • Teacher Model: GPT-2
  • Student Model: DistilGPT2

This alignment ensures the student model learns meaningful representations from the teacher while being more compact and efficient.

Differential Privacy Integration

Integrated Differentially Private Stochastic Gradient Descent (DP-SGD) into the training process. This provides mathematical guarantees about privacy:

  • Bounds the influence any single training example can have
  • Adds calibrated noise to gradients during training
  • Enables quantifiable privacy budgets (ε, δ)

Results

The approach enhances data confidentiality without significant loss in model performance, making it practical for real-world deployment where privacy is a concern.

Framework Re-engineering

In addition to the core research, I:

  • Re-engineered a text dataset distillation framework using Hugging Face models
  • Redesigned data-loading pipelines for efficiency and compatibility
  • Initiated full algorithm reimplementation to enable seamless integration with differential privacy mechanisms

Publications & Presentations

Research ongoing - details to be updated upon publication

Skills & Technologies

  • Models: GPT-2, DistilGPT2, Hugging Face Transformers
  • Privacy: Differential Privacy, DP-SGD, Privacy Accounting
  • Techniques: Knowledge Distillation, Layer Alignment
  • Tools: PyTorch, Hugging Face, Opacus