Attack and defense techniques in large language models: A survey and new perspectives

Area of Science:

• Artificial Intelligence
• Cybersecurity
• Natural Language Processing

Background:

• Large Language Models (LLMs) are integral to NLP but possess significant security and ethical vulnerabilities.
• The rapid advancement of LLMs necessitates a comprehensive understanding of their threat landscape.

Purpose of the Study:

• To systematically survey and categorize current attack and defense techniques targeting LLMs.
• To identify challenges and open issues in developing secure and resilient LLM systems.

Main Methods:

• Classification of LLM attacks into categories: adversarial prompt, optimized, model theft, and application-specific attacks.
• Analysis of defense strategies, including prevention-based and detection-based methods.
• Review of existing literature to synthesize the state-of-the-art in LLM security.

Main Results:

• Detailed mechanisms and implications of various LLM attack vectors are presented.
• Both prevention and detection-based defense strategies are analyzed, with their respective strengths and limitations.
• Identified challenges include adapting to evolving threats, balancing security with usability, and resource constraints.

Conclusions:

• Developing secure and resilient LLMs requires addressing adaptive scalability, robust adversarial attack detection, and generalized defense mechanisms.
• Ethical considerations and interdisciplinary collaboration are crucial for mitigating risks in real-world LLM applications.
• Ongoing research is needed to keep pace with the dynamic threat landscape and ensure AI safety.