Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EAGER: Edge-Aligned LLM Defense for Robust, Efficient, and Accurate Cybersecurity Question Answering

Onat Gungor, Roshan Sood, Jiasheng Zhou, Tajana Rosing

TL;DR

EAGER addresses the need for robust, accurate cybersecurity QA on edge devices by co-designing quantization-aware fine-tuning with domain-specific preference alignment. It integrates 4-bit quantized base weights (via NF4) with trainable low-rank adapters (QLoRA) and Direct Preference Optimization (DPO) trained on a self-generated cybersecurity preference dataset, enabling efficient edge deployment while preserving reasoning and safety. The framework achieves up to 7.3× reductions in attack success rate and up to 55% improvements in QA accuracy, with the lowest latency observed on Jetson Orin among evaluated defenses, demonstrating practical viability for SOC-like workflows at the edge. These results indicate that joint optimization of efficiency, robustness, and task accuracy through co-designed adaptation and domain-specific preferences can substantially improve lightweight LLM-based cybersecurity QA in resource-constrained environments.

Abstract

Large Language Models (LLMs) are highly effective for cybersecurity question answering (QA) but are difficult to deploy on edge devices due to their size. Quantization reduces memory and compute requirements but often degrades accuracy and increases vulnerability to adversarial attacks. We present EAGER, an edge-aligned defense framework that integrates parameter-efficient quantization with domain-specific preference alignment to jointly optimize efficiency, robustness, and accuracy. Unlike prior methods that address these aspects separately, EAGER leverages Quantized Low-Rank Adaptation (QLoRA) for low-cost fine-tuning and Direct Preference Optimization (DPO) on a self-constructed cybersecurity preference dataset, eliminating the need for human labels. Experiments show that EAGER reduces adversarial attack success rates by up to 7.3x and improves QA accuracy by up to 55% over state-of-the-art defenses, while achieving the lowest response latency on a Jetson Orin, demonstrating its practical edge deployment.

EAGER: Edge-Aligned LLM Defense for Robust, Efficient, and Accurate Cybersecurity Question Answering

TL;DR

EAGER addresses the need for robust, accurate cybersecurity QA on edge devices by co-designing quantization-aware fine-tuning with domain-specific preference alignment. It integrates 4-bit quantized base weights (via NF4) with trainable low-rank adapters (QLoRA) and Direct Preference Optimization (DPO) trained on a self-generated cybersecurity preference dataset, enabling efficient edge deployment while preserving reasoning and safety. The framework achieves up to 7.3× reductions in attack success rate and up to 55% improvements in QA accuracy, with the lowest latency observed on Jetson Orin among evaluated defenses, demonstrating practical viability for SOC-like workflows at the edge. These results indicate that joint optimization of efficiency, robustness, and task accuracy through co-designed adaptation and domain-specific preferences can substantially improve lightweight LLM-based cybersecurity QA in resource-constrained environments.

Abstract

Large Language Models (LLMs) are highly effective for cybersecurity question answering (QA) but are difficult to deploy on edge devices due to their size. Quantization reduces memory and compute requirements but often degrades accuracy and increases vulnerability to adversarial attacks. We present EAGER, an edge-aligned defense framework that integrates parameter-efficient quantization with domain-specific preference alignment to jointly optimize efficiency, robustness, and accuracy. Unlike prior methods that address these aspects separately, EAGER leverages Quantized Low-Rank Adaptation (QLoRA) for low-cost fine-tuning and Direct Preference Optimization (DPO) on a self-constructed cybersecurity preference dataset, eliminating the need for human labels. Experiments show that EAGER reduces adversarial attack success rates by up to 7.3x and improves QA accuracy by up to 55% over state-of-the-art defenses, while achieving the lowest response latency on a Jetson Orin, demonstrating its practical edge deployment.

Paper Structure

This paper contains 14 sections, 3 equations, 5 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. EAGER Framework
  4. Prompt Injection Attack Generation
  5. Prompt Injection Defense Formulation
  6. Cybersecurity Preference Dataset
  7. Direct Preference Optimization (DPO)
  8. Quantization-Aware Low-Rank Adaptation
  9. Co-Designed Preference Alignment
  10. Experimental Analysis
  11. Experimental Setup
  12. Baselines and State-of-the-Art Defenses
  13. Results
  14. Conclusion

Figures (5)

  • Figure 1: Comparison of existing defenses SecAlign chen2025meta, Q-resafe chenassessing, LlamaGuard inan2023llama, and our method (EAGER)
  • Figure 2: Overview of EAGER, our co-designed framework for cybersecurity QA. EAGER jointly integrates quantization-aware fine-tuning with domain-specific preference alignment to balance efficiency, robustness, and QA accuracy.
  • Figure 3: Comparison of EAGER and standard baselines.
  • Figure 4: Comparison of SOTA defense models SecAlign chen2025meta, LlamaGuard inan2023llama, Sandwich Defense learnprompting_sandwich, and Q-resafe chenassessing
  • Figure 5: SOTA latency comparison on Jetson Orin