Applicability of LLM in Assessing the Reliability of Members to Mitigate Internal Threats

As the damage caused by cyber attacks is becoming more serious, it is necessary to take countermeasures not only against external attacks but also against internal threats. There are various indicators of internal threats, such as accessibility to resources and the importance of those resources, but one of them is the level of reliability of members. For the reliability of members, access logs to resources, whether or not they have taken security training, their password settings, and their posts on Social Networking Service are considered as information sources. However, for reliability evaluation, these information sources need to be in a specific format. In addition, for information sources that can be evaluated from multiple aspects, processing is required for each aspect. To solve these problems, this paper investigates whether Large Language Model (LLM) performs well enough to evaluate the reliability of the members. We propose a framework that uses LLM to realize reliability assessment using internal organizational information sources. The framework enables members' reliability assessment in various aspects and allows the use of information sources even if they were written in common natural language. In order to verify the potential feasibility of the framework, we investigated the difference between the mathematical evaluation and the evaluation of the linguistic processing by LLM for two types of documents related to business evaluation written in numerical/natural language.