Benchmarking Out of the Box Open-Source LLMs for Malware Detection Based on API Calls Sequences

Benchmarking Out of the Box Open-Source LLMs for Malware Detection Based on API Calls Sequences

연구 분야: Safety

학회: International Conference on Intelligent Data Engineering and Automated Learning

The rise of generative AI has created several scenarios where older technologies (or human intervention) can be replaced by an agent that relies on LLMs. In this paper, we evaluate if an LLM is suited for malware detection and in what scenarios. For this task, we compare 4 open source models [LLama2-13B, Mistral, Mixtral and Mixtral-FP16] using a set of 20000 malware and 20000 benign files for which we provide behavioral information as a list of API calls sequences. The goal is to identify scenarios where these types of models can be successfully used for malware detection.