Master's Thesis: PEFT and Jailbreak Robustness
Jailbreak robustness of PEFT versus full fine-tuning on consumer-accessible LLMs
Do Efficient Adaptations Reduce Safety? Jailbreak Robustness of PEFT vs. Full Fine-Tuning on Consumer-Accessible LLMs
Master’s Thesis, Purdue University Fort Wayne, defended May 2026 Advisor: Dr. Jonathan Rusert; Committee: Dr. Anshuman Misra and Dr. Jay Johns
This thesis presents a systematic empirical study of how full fine-tuning, LoRA, QLoRA, and quantization affect jailbreak robustness in consumer-accessible language models. The study covers 25 model configurations, four jailbreak attacks, two safety benchmarks, three automated evaluators, and nearly 600 attack-success-rate measurements, calibrated against a 750-sample human annotation study.
Research Questions
Principal Findings
- Full fine-tuning, LoRA, and QLoRA all increased mean black-box attack success in the primary analysis.
- Attack type interacted significantly with adaptation strategy: behavior that improved against a white-box attack could worsen against black-box attacks.
- Evaluator choice materially changed measured safety; human validation favored the Llama-2 evaluator over GPT-4o-mini for precision on the annotated sample.
- Four-bit quantization of the base model preserved safety within the uncertainty of the study.
Methodology
- Empirical evaluation using PAIR, DeepInception, ArtPrompt, and AutoDAN
- Testing Gemma-2, Llama-3.1, Qwen-3, Phi-4, and Qwen-2.5 model families
- Reproducible FSDP and SLURM training and evaluation pipelines on Purdue Gilbreth
- Statistical analysis with bootstrap confidence intervals, hypothesis tests, and interaction analysis
Skills & Tools
Python, PyTorch, Transformers, PEFT, LoRA/QLoRA, Quantization, Adversarial Evaluation, LLM Safety