IMECA method of risk-based assessment and ensuring cybersecurity of Large Language Models

Abstract

Large Language Models (LLMs) help perform complex tasks that previously relied only on humans. Many different areas of human activity already use this technology or are actively exploring its capabilities with a view to future integration into work processes. In addition to the positive effects of their use, there are problems of uncertain and unexpected behavior, in particular, the generation of forbidden content. Given the expanding use of these models and their behavior, it is necessary to assess the level of security and further ensure the cybersecurity of this technology. The object of the study is the processes of ensuring cybersecurity for large language models. The article proposes a methodology for countering this threat by assessing the risks of such behavior and ensuring an acceptable level of cybersecurity for LLMs using the IMECA (Intrusion Modes Effects Criticality Analysis) technique of risk-based assessment. A set of countermeasures has been developed to increase the security level of LLMs, and procedures for their selection based on the criteria of maximum productivity and best rating using a countermeasure rating matrix are defined. An example of testing and ensuring the cybersecurity of a test language model is provided, the results of which show that the criticality level of cyber risks before the use of countermeasures is significantly decreased after using the most productive and highest-rated countermeasures, but threats with a high level of cyber risk criticality still exist. Directions for future research are proposed to deepen the procedure for evaluating and ensuring the security of LLMs, given the continuous development of these models and their protection mechanisms. The main result of this work is the combination of a model for ensuring the cybersecurity of LLMs and an improved method for analyzing the criticality of their vulnerabilities for further adaptation of the IMECA method of quantitative risk-based assessment and cybersecurity assurance for the field of LLMs.