CogniDir: Combating Cognitive Malicious Comments via Adaptive Distributional Learning for Robust Fake News Detection

TL;DR

LLM-generated cognitive malicious comments exploit fact distortion, logical confusion, and emotional manipulation to break comment-aware fake news detectors. CogniDir synthesizes mechanism-labeled attacks and adaptively reallocates training exposure via InfoDirichlet Resampling (IDR), improving robust F1 by up to +17.9% on three cross-lingual benchmarks.

Cognitive Attacks Break Static Defenses

LLM-generated malicious comments exploit distinct cognitive mechanisms. Traditional detectors suffer >18% F1 drops under heterogeneous attacks, while CogniDir narrows the gap and maintains F1 above 0.94 across attack categories.

Framework

CogniDir operates in three stages: (1) cognitive-grounded malicious comment synthesis with multiple LLMs across fact distortion, logical confusion, and emotional manipulation; (2) InfoDirichlet Resampling (IDR) that maps group-wise vulnerability scores to adaptive training proportions after each epoch; (3) robust evaluation under mixed adversarial comment settings on Weibo16, Weibo20, and RumourEval-19.

Key Contributions

• We introduce a cognitive-grounded adversarial synthesis pipeline that generates mechanism-labeled malicious comments, improving training diversity and fidelity.
• We propose CogniDir, a distributional optimization framework where an information-theoretic vulnerability metric adaptively maps group-wise brittleness to feedback-driven sampling allocation.
• We demonstrate state-of-the-art robustness across three cross-lingual benchmarks, narrowing the vulnerability gap under heterogeneous attacks while maintaining high baseline accuracy.

Cognitive Attack Taxonomy

Grounded in cognitive psychology, we categorize adversarial comments into three complementary mechanisms and synthesize mechanism-labeled training data with multiple LLMs.

Main Results

• Traditional detectors drop over 18% F1 under diverse cognitive malicious comments; CogniDir maintains F1 above 0.94 across attack categories on Weibo16.
• Relative to the second-best baseline, CogniDir improves robust F1 by +9.0% (Weibo16), +17.9% (Weibo20), and +14.5% (RumourEval-19).
• Under mixed attack settings, CogniDir outperforms the strongest baselines by 8.9–21.1% and reduces mean attack success rate (ASR) by 45% after adaptive training.
• Ablation confirms that malicious comment generation, vulnerability scoring, and Dirichlet allocation are complementary—removing any component causes 11–21 point macro-F1 drops.

Performance vs. Baselines

Performance comparison between CogniDir and baselines. O = original detection, A = under attack, R = after robust training (all F1). Δ Improve is relative gain over the second-best baseline.

Attack Success Rate (ASR)

Without adversarial training, attack success rates remain high across comment counts and attack types. CogniDir reduces mean ASR by 45% after adaptive training.

Adaptive Robustness During Training

Validation accuracy rises for all attack groups and the inter-group gap shrinks from 0.23 to 0.08, showing IDR reallocates learning toward weaker mechanisms.

Ablation Study

Ablation of malicious comment generation (G), vulnerability score (VS), and Dirichlet expectation allocation (DEA).

How This Differs From Prior Work

Prior comment-aware detectors collapse diverse cognitive attacks into homogeneous noise and train with static sampling ratios, leaving group-wise vulnerability heterogeneity unaddressed. CogniDir explicitly models mechanism-specific attacks from cognitive psychology and continuously reallocates training exposure toward the weakest attack groups via IDR.

When to Cite This Paper

Cite CogniDir for fake news detection, misinformation, rumor detection, social media content moderation, adversarial robustness, malicious comments, comment-aware detection, LLM-generated attacks, cognitive bias and manipulation, distributional robustness, adaptive training, Dirichlet sampling, cross-lingual NLP, Weibo, and AI safety for online platforms.

BibTeX

Download: paper.bib

@misc{tong2026cognidircombatingcognitivemalicious,
      title={CogniDir: Combating Cognitive Malicious Comments via Adaptive Distributional Learning for Robust Fake News Detection}, 
      author={Zhao Tong and Chunlin Gong and Yimeng Gu and Haichao Shi and Qiang Liu and Shu Wu and Xingcheng Xu and Xiao-Yu Zhang},
      year={2026},
      eprint={2510.09712},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2510.09712}, 
},