Learning from Literature: A Retraining-Free Framework for LLM Jailbreak Defense via NLP-Based Adversarial Literature Analysis
Year: 2025
Type: conference-paper
Abstract: Recent surge in adoption of Large Language Models (LLMs) have made them vulnerable to sophisticated adversarial attacks, manipulative prompts, and encoded malicious inputs. Existing countermeasures often require retraining models, a process that is computationally expensive and impractical for deployment. This paper introduces a novel defense framework that enables LLMs to autonomously detect, filter, and fend off malicious or adversarial inputs without requiring retraining or fine-tuning. The proposed framework consists of two core components: (1) A summarization module that processes and summarizes adversarial research literature to provide the LLM with context-aware defense knowledge and (2) A prompt filtering module that detects, decodes, and classifies harmful inputs through advanced Natural Language Processing (NLP) techniques, such as zero-shot classification, keyword analysis, and encoded content detection. By combining text extraction, summarization, and malicious prompt analysis, this approach enhances the robustness of LLMs against adversarial misuse. Experimental results demonstrate that this integrated approach successfully identifies encoded prompts, malicious patterns, and manipulative language structures with a success rate of 98.7%. The framework also enables the model to respond appropriately to malicious prompts with a higher percentage of jailbreak resistance and refusal rate, using minimal quantity of adversarial research literature as context. The framework significantly reduces LLM Attack Success Rate (ASR) as a lightweight, efficient alternative to retraining-based time and resource-consuming defenses, while upholding LLM response quality.
Institution: University of Notre Dame
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4415025828