EMNLP 2026 Main Conference, BudapestMain2026

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

Zhao Tong*1,2, Chunlin Gong*3, Yimeng Gu4, Haichao Shi1, Qiang Liu5, Shu Wu†5, Xingcheng Xu6, Xiao-Yu Zhang†1

  1. 1Institute of Information Engineering, Chinese Academy of Sciences
  2. 2School of Cyber Security, University of Chinese Academy of Sciences
  3. 3University of Minnesota
  4. 4Queen Mary University of London
  5. 5New Laboratory of Pattern Recognition (NLPR), Institute of Automation, Chinese Academy of Sciences
  6. 6Shanghai AI Laboratory

*Equal contribution: Zhao Tong, Chunlin Gong · †Corresponding author: Shu Wu, Xiao-Yu Zhang

+17.9%Best F1 Gain (Weibo20)
0.947Weibo20 Robust Macro-F1
45%Mean ASR Reduction
3Benchmarks

TL;DR

CogniDir strengthens comment-aware fake news detection through mechanism-specific malicious comment generation and feedback-driven training mixtures. InfoDirichlet Resampling adapts exposure across attack groups on Weibo16, Weibo20, and RumourEval-19.

Research context

Comment-aware fake news detectors combine news content with social responses. LLM-generated malicious comments can manipulate this context through fact distortion, logical confusion, and emotional manipulation. CogniDir studies these mechanisms as distinct sources of detector vulnerability.

The framework connects adversarial social-context evaluation with adaptive training-data allocation. It scores vulnerability across attack groups and uses a Dirichlet-mean allocation to update training exposure. Experiments on Weibo16, Weibo20, and RumourEval-19 evaluate robustness across attack mechanisms and comment settings.

Research noteRobust Fake News Detection under Adversarial Social Context

Cognitive Attacks Break Static Defenses

On Weibo16, prior methods drop sharply under diverse malicious comments. CogniDir maintains high robust F1, staying near 0.94 across attack categories.

Group-wise F1 comparison under cognitive malicious comments
Figure 1. Group-wise F1 on Weibo16: prior methods drop sharply under diverse malicious comments; CogniDir maintains high robust F1. · Vector PDF

Framework

CogniDir works in three stages: mechanism-specific malicious comment generation, InfoDirichlet Resampling that updates the training mixture from group vulnerability after each epoch, and robust evaluation under mixed adversarial comments. The mixture uses the Dirichlet mean, a smooth closed-form allocation that tracks the weakest attack group.

CogniDir framework overview
Figure 2. Cognitive malicious comment synthesis, InfoDirichlet Resampling (IDR), and robust evaluation under mixed attack settings. · Vector PDF

Key Contributions

  • A cognitive-grounded synthesis pipeline that produces mechanism-labeled comments for fact distortion, logical confusion, and emotional manipulation.
  • InfoDirichlet Resampling, which turns group-wise vulnerability into an adaptive training mixture through the Dirichlet mean.
  • Strong robustness on Weibo16, Weibo20, and RumourEval-19, with up to +17.9% F1 and a 45% reduction in mean ASR.

Cognitive Attack Taxonomy

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

FACT

Fact Distortion

Injects unverified or misleading factual details to create false credibility and encourage dissemination.

LOGIC

Logical Confusion

Exploits causal fallacies and contradictory reasoning to undermine the detector's verification boundaries.

EMOTION

Emotional Manipulation

Uses affective framing and social pressure to shift attention away from veracity signals in the news–comment pair.

Main Results

  • On Weibo16, CogniDir keeps F1 near 0.94 across attack categories, while prior detectors drop sharply under diverse malicious comments.
  • Robust F1 improves by +9.0% on Weibo16, +17.9% on Weibo20 (Macro-F1 0.947), and +14.5% on RumourEval-19 over the strongest baselines.
  • Adaptive training reduces mean ASR by 45%, down to 0.116, across attack types and comment counts.
  • Malicious-comment generation, vulnerability scoring, and Dirichlet expectation allocation each contribute to the final robustness.

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.

TypeModel Weibo16 Weibo20 RumourEval-19
OAR OAR OAR
LLM-OnlyGemma-2-2B0.3220.2910.2740.3450.2920.3260.2940.2380.275
Mistral-7B0.6860.3340.4950.6540.3150.3750.4680.1840.252
Llama-3-8B0.5980.3240.4080.5820.3630.4080.3770.2190.276
Qwen2.5-32B0.8310.2820.7160.8230.3670.5410.5530.3430.412
Deep-Learning BaseddEFEND0.9020.7270.7450.8840.5570.5850.6480.4070.687
Dual-CAN0.8950.3660.5780.8720.4360.5800.6280.3770.546
GenFEND0.9150.5780.8670.8920.6400.7610.6620.4560.606
L-Defense0.8820.7060.7590.8600.6660.7020.6100.3430.709
ARG0.8840.6190.8200.8120.5760.8030.7120.4400.609
CogniDir0.9550.7920.9450.9360.7280.9470.8220.5520.812
Δ Improve+4.4%+8.9%+9.0%+4.9%+9.3%+17.9%+15.5%+21.1%+14.5%

Attack Success Rate (ASR)

Adaptive training cuts mean attack success rate by 45%, bringing it down to 0.116 across attack types and comment counts.

ASR before robust training
Figure 3a. ASR before robust training under varying attack types and comment counts. · Vector PDF
ASR after CogniDir training
Figure 3b. ASR after robust training — substantial reduction across attack types. · Vector PDF

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.

Validation accuracy convergence across attack groups
Figure 4. Group-wise validation accuracy during training — Fact Distortion, Logical Confusion, and Emotional Manipulation. · Vector PDF

Ablation Study

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

DatasetMethodMacro-F1Acc.F1-realF1-fake
Weibo16CogniDir-G0.7920.8110.8300.754
CogniDir-VS0.8210.8360.8620.780
CogniDir-DEA0.8550.8720.8950.815
CogniDir0.9450.9470.9570.933
Weibo20CogniDir-G0.7280.7420.7700.686
CogniDir-VS0.7920.8070.8250.759
CogniDir-DEA0.8340.8480.8720.796
CogniDir0.9470.9470.9510.943
RumourEval-19CogniDir-G0.5520.5740.6200.484
CogniDir-VS0.6130.6280.6620.564
CogniDir-DEA0.6920.7070.7480.636
CogniDir0.8120.8360.8830.741

Allocation

CogniDir assigns training exposure to distinct attack mechanisms and updates the mixture from group-wise vulnerability feedback. InfoDirichlet Resampling maps the scores to positive Dirichlet parameters and uses the normalized mean as the next training mixture.

Research findings

  • Mechanism-specific attacks. Fact distortion, logical confusion, and emotional manipulation define distinct adversarial comment groups.
  • Feedback-driven training. Group-wise vulnerability scores guide the next training mixture through InfoDirichlet Resampling.
  • Robustness evaluation. Experiments on Weibo16, Weibo20, and RumourEval-19 measure detection performance and attack success across the reported settings.

Code and resources

  • Training — Run the detector training entry point.
  • Evaluation — Run model evaluation, including group-wise evaluation.
  • Data format — Read the data preparation and format documentation.
  • BERT setup — Read the pretrained BERT model setup instructions.
  • Generation model setup — Read the comment-generation model setup instructions.

Questions

How do social comments affect fake news detection?

Comment-aware detectors use social responses together with news content. Manipulated comments can alter this evidence and affect the detector's predictions. CogniDir evaluates this behavior under LLM-generated cognitive malicious comments.

Which cognitive attack mechanisms are studied?

The study evaluates fact distortion, logical confusion, and emotional manipulation as separate comment attack groups. Group-wise results characterize detector vulnerability across these mechanisms.

How does CogniDir adapt adversarial training?

After each epoch, the framework scores attack-group vulnerability using accuracy and probabilistic confidence. InfoDirichlet Resampling converts these scores into Dirichlet parameters and uses the Dirichlet mean to set the next training mixture.

How does group-wise evaluation guide training exposure?

Group-wise measurements track how the detector responds to each attack mechanism. The allocation gives vulnerable groups additional exposure and updates the mixture as the detector learns.

Which datasets and languages are evaluated?

The experiments use Weibo16, Weibo20, and RumourEval-19, covering Chinese and English comment-based fake news detection. The evaluation reports detection performance across attack mechanisms and comment settings.

Which components are available in the repository?

The repository provides training and evaluation entry points, model and sampling modules, and data-format documentation. The project resources link to the training script, evaluation script, and setup instructions.

Publication history

Earlier arXiv versions of this work were titled "Group-Adaptive Adversarial Learning for Robust Fake News Detection Against Malicious Comments." The published EMNLP 2026 version is titled "CogniDir: Combating Cognitive Malicious Comments via Adaptive Distributional Learning for Robust Fake News Detection" (arXiv:2510.09712).

arXiv version history

BibTeX

Official EMNLP BibTeX. Download: paper.bib

@inproceedings{tong-etal-2026-cognidir,
 title = "{C}ogni{D}ir: Combating Cognitive Malicious Comments via Adaptive Distributional Learning for Robust Fake News Detection",
 author = "Tong, Zhao and Gong, Chunlin and Gu, Yimeng and Shi, Haichao and Liu, Qiang and Wu, Shu and Xu, Xingcheng and Zhang, Xiao-Yu",
 editor = "Lee, Hung-Yi and Dev, Sunipa and Li, Jessy and Elliott, Desmond",
 booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
 month = oct,
 year = "2026",
 address = "Budapest, Hungary",
 publisher = "Association for Computational Linguistics",
 url = "https://aclanthology.org/2026.emnlp-main.655/",
 pages = "15050--15067",
 ISBN = "979-8-89176-459-0"
}