Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RADAR: Retrieval-Augmented Detector with Adversarial Refinement for Robust Fake News Detection

Song-Duo Ma, Yi-Hung Liu, Hsin-Yu Lin, Pin-Yu Chen, Hong-Yan Huang, Shau-Yung Hsu, Yun-Nung Chen

TL;DR

RADAR introduces a retrieval-augmented detector with Verbal Adversarial Feedback to robustly detect fake news under adaptive attacks. By pairing generator-side retrieval ( realism priors ) with detector-side retrieval ( external evidence for verification), and replacing scalar rewards with structured VAF, RADAR enables stable co-evolution of attack and defense. The detector is a lightweight DeBERTa-based encoder trained with cross-entropy and guided by VAF, while the generator is a LoRA-finetuned Qwen-4B model that rewrites real articles into convincing fakes. Empirical results on AdvFake-News-Please show a ROC-AUC of 86.98% with full dual-RAG, with detector retrieval and VAF contributing the largest gains and outperforming general-purpose LLM baselines under retrieval settings.

Abstract

To efficiently combat the spread of LLM-generated misinformation, we present RADAR, a retrieval-augmented detector with adversarial refinement for robust fake news detection. Our approach employs a generator that rewrites real articles with factual perturbations, paired with a lightweight detector that verifies claims using dense passage retrieval. To enable effective co-evolution, we introduce verbal adversarial feedback (VAF). Rather than relying on scalar rewards, VAF issues structured natural-language critiques; these guide the generator toward more sophisticated evasion attempts, compelling the detector to adapt and improve. On a fake news detection benchmark, RADAR achieves 86.98% ROC-AUC, significantly outperforming general-purpose LLMs with retrieval. Ablation studies confirm that detector-side retrieval yields the largest gains, while VAF and few-shot demonstrations provide critical signals for robust training.

RADAR: Retrieval-Augmented Detector with Adversarial Refinement for Robust Fake News Detection

TL;DR

RADAR introduces a retrieval-augmented detector with Verbal Adversarial Feedback to robustly detect fake news under adaptive attacks. By pairing generator-side retrieval ( realism priors ) with detector-side retrieval ( external evidence for verification), and replacing scalar rewards with structured VAF, RADAR enables stable co-evolution of attack and defense. The detector is a lightweight DeBERTa-based encoder trained with cross-entropy and guided by VAF, while the generator is a LoRA-finetuned Qwen-4B model that rewrites real articles into convincing fakes. Empirical results on AdvFake-News-Please show a ROC-AUC of 86.98% with full dual-RAG, with detector retrieval and VAF contributing the largest gains and outperforming general-purpose LLM baselines under retrieval settings.

Abstract

To efficiently combat the spread of LLM-generated misinformation, we present RADAR, a retrieval-augmented detector with adversarial refinement for robust fake news detection. Our approach employs a generator that rewrites real articles with factual perturbations, paired with a lightweight detector that verifies claims using dense passage retrieval. To enable effective co-evolution, we introduce verbal adversarial feedback (VAF). Rather than relying on scalar rewards, VAF issues structured natural-language critiques; these guide the generator toward more sophisticated evasion attempts, compelling the detector to adapt and improve. On a fake news detection benchmark, RADAR achieves 86.98% ROC-AUC, significantly outperforming general-purpose LLMs with retrieval. Ablation studies confirm that detector-side retrieval yields the largest gains, while VAF and few-shot demonstrations provide critical signals for robust training.
Paper Structure (41 sections, 2 equations, 3 figures, 5 tables, 1 algorithm)

This paper contains 41 sections, 2 equations, 3 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Fake News Detection
  4. Retrieval-Augmented Evidence for Misinformation Tasks
  5. Adversarial Sample Generation and Text GANs
  6. Natural Language Feedback for Optimization
  7. Methodology
  8. Task Definition
  9. RADAR Framework
  10. Retrieval Module
  11. Generator
  12. Detector
  13. Verbal Adversarial Feedback
  14. Suspicious Token Identification
  15. Detection Reasons
  16. ...and 26 more sections

Figures (3)

  • Figure 1: The x-axis indicates generator-side retrieval, and the y-axis indicates detector-side retrieval. Markers distinguish adversarial methods from non-adversarial ones. Our RADAR is an adversarial, dual-side design.
  • Figure 2: Overview of RADAR for robust fake news detection under adaptive attacks. Retrieval augments both sides using a real-news database: generator-side RAG strengthens the attacker by providing realistic journalistic priors for fact-altered rewriting, while detector-side RAG supplies external evidence for evidence-aware detection. The detector also outputs VAF signals to guide iterative attack refinement, enabling continual co-evolution.
  • Figure 3: Training dynamics of detector ROC-AUC (%) over 6 rounds. The G$+$/D$+$ configuration consistently achieves the strongest performance, demonstrating the benefit of dual-side retrieval.