PEFT and Jailbreak Robustness

Master's thesis on robustness-efficiency trade-offs in instruction-tuned LLMs

Robustness-Efficiency Trade-offs in PEFT-based Instruction-Tuned LLMs

Master’s Thesis, Purdue University Fort Wayne
Advisor: Dr. Jonathan Rusert

This thesis empirically studies how parameter-efficient fine-tuning (PEFT) techniques like LoRA and QLoRA, combined with quantization, affect the jailbreak robustness of consumer-sized language models. The core research question is whether the same techniques that make models cheaper and easier to deploy can quietly erode safety margins under realistic adversarial prompting.

Research Questions

  • How do PEFT methods (LoRA, QLoRA) interact with quantization to influence model robustness against jailbreak attacks?
  • What are the trade-offs between computational efficiency and adversarial robustness in instruction-tuned models?
  • Can we identify systematic patterns in how efficiency optimizations impact safety alignment?

Methodology

  • Empirical evaluation using modern black-box jailbreak attack families
  • Testing on consumer-accessible model sizes (7B-13B parameters)
  • Building evaluation harnesses for reproducible robustness experiments
  • Comprehensive analysis of failure modes under different PEFT and quantization configurations

Target Venue

EMNLP 2026

Skills & Tools

Python, PyTorch, Transformers, PEFT, LoRA/QLoRA, Quantization, Adversarial Evaluation, LLM Safety