SECURITY THREATS AND PROTECTION STRATEGIES IN LARGE LANGUAGE MODELS

Large Language Models (LMMs) have become an indispensable part of modern technologywith their groundbreaking capabilities in natural language processing. These models have awide range of applications, from virtual assistants to code development, but they also poseserious security and privacy threats. This study examines in detail the major security threats toLLMs: backdoor attacks, model inversion attacks, membership inference attacks, and datapoisoning. Especially in sensitive data processing areas such as healthcare, finance and law,risks such as leakage of personal information and malicious content generation are of greatconcern. This study addresses various protection strategies against these threats such as dataencryption, access control, attacker training and differential privacy. It also emphasizes thatinternational legal regulations such as GDPR, HIPAA, CCPA and the EU Artificial IntelligenceAct (AI Act) provide a critical framework for the responsible and ethical development and useof LLMs. As a result, it is stated that ensuring the security of LLMs will be possible not onlywith technical measures, but also with legal compliance, ethical principles and interdisciplinarycollaboration.