Realistic Log Generation through Large Language Models: Bridging Data Gaps in Log Analysis

Abstract

System logs are very useful for monitoring, debugging and diagnosing failures of large systems but they tend to be sensitive and thus not easy to share and make use of in a research. We propose a solution to this problem: to produce synthetic logs information of sufficiently good quality such that the structure and the statistical distribution of the real logs and privacy information are preserved. The strategy has three steps. The former part is masking the sensitive material on the logs by using a multi-pass algorithm of anonymization that generates a dictionary of sensitive words and substitutes them with generic names. Second, using the anonymized logs synthetic logs are generated using large language models (LLMs) and to ensure continuity between chunks summaries of previously generated synthetic log chunks are used and reflective verification loops also. Lastly, the produced logs are fed through a multi-agent debating model which is automated and functions using an LLM pipeline to achieve the structural and contextual correctness. With the help of this model, organizations can generate viable, interchangeable logs without revealing information that is confidential that can be used to aid in the completion of tasks such as log enquiry, failure enquiry, and root-cause research.

Keywords

Article information