Defending LLMs Against Adversarial Prompts: A Gradient-Correlation Approach with Graph-Based Parameter Analysis

ACL ARR 2025 May Submission8077 Authors

Abstract: Large Language Models (LLMs) face covert threats from toxic prompts, and existing detection methods often require substantial data and are inefficient. Cur-rent gradient-based approaches primarily focus on individual parameter compari-sons, limiting their effectiveness against sophisticated toxicity. To address this, we propose GradMesh, which integrates Eu-clidean distance metrics for gradient magnitudes with direction similarity analysis. We also employ Graph Neural Networks (GNN) to model relationships among parameters, enhancing detection accuracy by clustering correlated parame-ters. Additionally, we generate diverse toxic reference samples using the target LLM to improve reliability. Experiments on benchmark datasets ToxicChat and XSTest show that GradMesh outperforms existing methods across all evaluation metrics.

Paper Type: Long

Research Area: Ethics, Bias, and Fairness

Research Area Keywords: model bias/unfairness mitigation, reflections and critiques

Contribution Types: NLP engineering experiment

Languages Studied: Chinese, English

Submission Number: 8077