UC Santa Barbara

The rapid advancement of powerful Large Language Models (LLMs), such as ChatGPT and Llama, has revolutionized the world by bringing new creative possibilities and enhancing productivity. However, these advancements also pose significant challenges and risks, including the potential for misuse in the form of fake news, academic dishonesty, intellectual property infringements, and privacy leaks. In response to these concerns, this thesis explores approaches to promoting the responsible use of LLMs from both theoretical and empirical perspectives.

Three key approaches are presented: (1) Detecting AI-generated Text via Watermarking: We propose a robust and high-quality watermarking method called Unigram-Watermark and introduce a rigorous theoretical framework to quantify the effectiveness and robustness of LLM watermarks. Furthermore, we propose PF-Watermark, which achieves the best balance of high detection accuracy and low perplexity. (2) Protecting the Intellectual Property of LLMs: We safeguard the intellectual property of LLMs through novel watermarking techniques designed to prevent model-stealing attacks in both text classification and text generation tasks. (3) Privacy-Preserving LLMs: We employ Confidential Redacted Training (CRT) to train and fine-tune language generation models while protecting sensitive information. In summary, we propose a suite of algorithms and solutions to address LLMs' trending safety, security, and privacy concerns. We hope our studies provide valuable insights for researchers to explore exciting future research solutions that promote responsible AI development and deployment.