Breach in the Shield: Unveiling the Vulnerabilities of Large Language Models

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Dai, Runpeng, Yang, Run, Zhou, Fan, Zhu, Hongtu
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866908435989135360
author Dai, Runpeng
Yang, Run
Zhou, Fan
Zhu, Hongtu
author_facet Dai, Runpeng
Yang, Run
Zhou, Fan
Zhu, Hongtu
contents Large Language Models (LLMs) and Vision-Language Models (VLMs) have achieved impressive performance across a wide range of tasks, yet they remain vulnerable to carefully crafted perturbations. In this study, we seek to pinpoint the sources of this fragility by identifying parameters and input dimensions (pixels or token embeddings) that are susceptible to such perturbations. To this end, we propose a stability measure called \textbf{FI}, \textbf{F}irst order local \textbf{I}nfluence, which is rooted in information geometry and quantifies the sensitivity of individual parameter and input dimensions. Our extensive analysis across LLMs and VLMs (from 1.5B to 13B parameters) reveals that: (I) A small subset of parameters or input dimensions with high FI values disproportionately contribute to model brittleness. (II) Mitigating the influence of these vulnerable parameters during model merging leads to improved performance.
format Preprint
id arxiv_https___arxiv_org_abs_2504_03714
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Breach in the Shield: Unveiling the Vulnerabilities of Large Language Models
Dai, Runpeng
Yang, Run
Zhou, Fan
Zhu, Hongtu
Machine Learning
Artificial Intelligence
Computation and Language
Large Language Models (LLMs) and Vision-Language Models (VLMs) have achieved impressive performance across a wide range of tasks, yet they remain vulnerable to carefully crafted perturbations. In this study, we seek to pinpoint the sources of this fragility by identifying parameters and input dimensions (pixels or token embeddings) that are susceptible to such perturbations. To this end, we propose a stability measure called \textbf{FI}, \textbf{F}irst order local \textbf{I}nfluence, which is rooted in information geometry and quantifies the sensitivity of individual parameter and input dimensions. Our extensive analysis across LLMs and VLMs (from 1.5B to 13B parameters) reveals that: (I) A small subset of parameters or input dimensions with high FI values disproportionately contribute to model brittleness. (II) Mitigating the influence of these vulnerable parameters during model merging leads to improved performance.
title Breach in the Shield: Unveiling the Vulnerabilities of Large Language Models
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2504.03714