Addressing Security Vulnerabilities Caused by Fine-Tuning in Large Language Models

Abstract

The introduction of large language models into our lives is revolutionary and has enabled widespread use of artificial intelligence and easy access to end users. With this development, large language models have begun to be actively used in many different sectors and have become an important part of daily life. However, the adaptation method implemented for large language models to gain competence in sector-specific usage areas weakens the original security measures of large language models and causes new security vulnerabilities to emerge. In particular, these adaptation stages create new attack surfaces that malicious people can exploit and increase concerns about the security of the models. The main reason for this is that the model is exposed to malicious patterns, either directly with malicious data or indirectly with biased datasets, and therefore the security alignment shifts. This study presents a solution proposal to reinforce this weakened security by revealing the weakened security of large language models through adaptation. In this context, a scoring method has been presented for the security of large language models against post-adaptation input manipulation attacks, and then the weakened security has been strengthened with the proposed security tightening method. The distinguishing feature of the study is that, unlike existing methods, an additional adaptation process has been applied to increase the security of large language models adapted to provide better answers in the Turkish language.

Ähnliche Arbeiten