A Survey on Knowledge-driven Autonomous Driving

PhD Thesis Proposal Defense


Title: "A Survey on Knowledge-driven Autonomous Driving"

by

Mr. Haoran LI


Abstract:

The rise of transformer models has significantly advanced machine learning 
models. Large language models (LLMs), trained on massive data and supported by 
extensive computational resources, unify the conventional Natural Language 
Processing (NLP) paradigm and can effectively handle a variety of downstream 
tasks by integrating these tasks into generative workflows. For real-world 
impacts, LLMs have revolutionized accessibility and usability for researchers, 
developers and users. Consequently, powerful LLMs have empowered new 
possibilities across various fields, including agents, smart assistants, 
chatbots, and more.

However, the widespread availability and accessibility of these models also 
introduce potential risks, including malicious use and privacy concerns. The 
free-form generation pipelines that make LLMs valuable can also be misused to 
compromise privacy or for harmful purposes. Even though heavy efforts have 
already been made to enhance LLMs' trustworthiness to target LLMs' safety and 
privacy issues, new attacks are frequently proposed to bypass existing defense 
mechanisms and comprise LLMs for malicious uses. Therefore, a consistent and 
adversarial game exists between malicious attackers and defenders regarding 
LLMs' trustworthiness, leaving significant challenges undiscovered.

To fully study LLMs' trustworthiness issues, we identify new attacks in terms 
of information leakage, improve defense mechanisms to address various attacks 
and empirically evaluate the attacks' outcomes with and without defenses. For 
identified attacks, we target information leakage issues on vector databases to 
study embeddings' information leakage. Beyond information leakage on 
embeddings, we also demonstrate how to compromise LLMs with jailbreaking 
prompts. After discussing attacks, we present new defenses to prevent 
information leakage on embeddings. Lastly, we implement a benchmark to evaluate 
attacks' performance with and without defenses empirically. Extensive 
experiments are conducted to support the effectiveness of our discovered 
attacks and defenses. Our evaluation benchmark's results reveal the gap between 
attacks' and defenses' assumptions.


Date:                   Tuesday, 16 July 2024

Time:                   10:00am - 12:00noon

Venue:                  Room 5510
                        Lifts 25/26

Committee Members:      Dr. Yangqiu Song (Supervisor)
                        Dr. Binhang Yuan (Chairperson)
                        Dr. Junxian He
                        Dr. Dongdong She