Security Threats in the Inference Phase of Large Language Models

Abstract

In recent years, Large Language Models (LLMs) have achieved remarkable advancements in natural language understanding, contextual modeling, and human-like reasoning. However, these capabilities have also unveiled a spectrum of critical security vulnerabilities, including logical inconsistencies during inference, linguistic and semantic disruptions, and entrenched cognitive biases. Such deficiencies can cause LLMs to produce inaccurate, misleading, or even harmful outputs, posing significant risks to high-stakes applications such as decision support systems, legal consultation, and medical diagnostics. This paper presents a systematic analysis of security threats arising during the inference phase of LLMs. By examining the inference pipeline across its distinct stages and the varying levels of abstraction at which threats manifest, we propose a comprehensive taxonomy of potential security risk factors. Furthermore, we conduct targeted testing and validation to assess the impact of these threats. Our study aims to provide both theoretical insights and practical guidance for the development and deployment of safer and more reliable LLM systems.

Ähnliche Arbeiten