Large Language Model-Based Reward Design for Deep Reinforcement Learning-Driven Autonomous Cyber Defense

The emergence of Large Language Model (LLM)-driven multi-agent systems has transformed software development, allowing users with minimal technical skills to create applications through natural language inputs. However, this innovation also raises significant security concerns, particularly through scenarios where malicious users exploit benign agents or vice versa. The introduction of the Implicit Malicious Behavior Injection Attack (IMBIA) highlights these vulnerabilities, with alarming success rates in various frameworks.

A new evaluation metric called SMILE has been introduced to enhance the assessment of question-answering systems by integrating both lexical exactness and semantic understanding. This metric aims to address the limitations of traditional methods that rely heavily on n-gram similarity, which often overlook deeper semantic meanings. SMILE combines sentence-level and keyword-level evaluations to provide a more comprehensive assessment of responses.