Adversarial Attacks and Defenses on Large Language Models: A Systematic Review

The rapid advancement of large language models (LLMs) presents both transformative opportunities and significant security challenges, primarily due to their susceptibility to adversarial attacks-crafted inputs designed to deceive them-which pose considerable risks to their reliability and security. This survey comprehensively reviews these textual adversarial attacks, categorizing them into sentence-level, wordlevel, and character-level manipulations, while also exploring multi-level approaches that combine these techniques. To counter such threats, the survey examines a range of defense strategies, including detection methods, adversarial training, semantic analysis, general system enhancements, and approaches to certified robustness for provable resilience. Despite notable progress in understanding and mitigating these vulnerabilities, significant challenges remain regarding real-world applicability, scalability, generalization across different models and attack types, and the establishment of standardized benchmarking. By synthesizing recent research in this dynamic field, this survey aims to guide the development of more secure and robust language models capable of withstanding the continuously evolving landscape of adversarial threats.