BanFakeMM: A Multimodal Bangla Fake News Detection Benchmark with Progressive Adversarial Attacks
Description
BanFakeMM is a large-scale, detector-guided multimodal Bengali fake news benchmark comprising 7,040 authentic news articles (with images) collected from eight Bangla newspapers across eleven topical categories, together with five progressively challenging adversarial datasets constructed via a five-round, detector-in-the-loop framework using Gemini 2.5 Flash Lite and Imagen 3/4. Each benchmark dataset used for detector training/evaluation is balanced with 1,000 real and 1,000 fake samples (2,000 total), split 80/10/10 into train/val/test: - AdversarialGen (1,000 fake sampled from a larger pool of 3,500 generated articles): fully LLM-generated fake news (text + image) across four categories (clickbait, hoax, propaganda, misleading) - LinguisticAttack (1,000 fake): real news perturbed via antonym, quantifier, and mixed mutations while preserving style - AdversarialEdit (1,000 fake): coordinated text-image manipulation with localized image editing - RandomSwap (1,000 fake): authentic text paired with a randomly swapped image from an unrelated article - SemanticSwap (1,000 fake): authentic text paired with a swapped image from a semantically similar (but factually distinct) article Every sample includes an image and a CSV row with schema (article_id, source, category, title, content, published_date, image_path, label, fake_type). This benchmark supports research on adversarially robust multimodal fake news detection, cross-modal consistency modeling, image forensics, and low-resource multilingual NLP. It was used to train/evaluate five detection architectures (BanglaBERT, Feature Late Fusion, TACMA, CMCF, and CMCGF) described in the accompanying paper "A Progressive Adversarial Benchmark and Cross-Modal Consistency Guided Detection for Bengali Multimodal Fake News Detection."
Files
Institutions
- Chittagong University of Engineering & TechnologyChittagong, Chittagong