Mitigating Attacks on LLMs with Privacy and Adversarial Training

Abstract

Mitigating Attacks on LLMs with Privacy and Adversarial TrainingOlasunkanmi Sodunke School of Engineering Technology, Purdue University The rapid adoption of Large Language Models (LLMs) such as BERT & GPT-3 has intensified privacy concerns, especially with vulnerabilities like model inversion and membership inference attacks that can exploit sensitive information from model training data. This research, led by myself, under the guidance of Reham Nour, School of Engineering Technology, Purdue University Indianapolis, explores the efficacy of integrating differential privacy and adversarial training to counter these threats while maintaining the utility and performance of LLMs. Differential privacy safeguards individual data points by injecting controlled noise, whereas adversarial training increases robustness by preparing models against adversarial examples.The study specifically examines the interaction between differential privacy and adversarial training to understand their combined impact on security and model performance. We employ a range of datasets, including SST-2, TextFooler, MIMIC-III, and AdvGLUE, to assess these mechanisms across various applications such as sentiment analysis and healthcare data privacy. Our methodology includes the use of Generalized Linear Models (GLM) to analyze the trade-offs between privacy protection and accuracy. Evaluation metrics will focus on performance (accuracy, F1-score) and privacy protection (privacy leakage metrics).This research aims to contribute significantly to the field of LLM security by providing insights into how these dual mechanisms can be synergistically applied to mitigate privacy risks effectively. The findings are intended to have practical implications for sectors dealing with sensitive data, enhancing the security posture of LLMs in both academic and industry settings.