BSM-DND: Bias and Sensitivity-Aware Multilingual Deepfake News Detection Using Bloom Filters and Recurrent Feature Elimination

Abstract

The rapid spread of deepfake news across digital media platforms poses a growing threat to societal trust, public safety, and global security. With the increasing realism of AI-generated content, distinguishing between authentic and fabricated information has become increasingly challenging, especially in low-resource and multilingual content. In this work, we propose a novel bias and sensitivity-aware multilingual deepfake news detection framework, BSM-DND, which classifies news articles into three categories: Authentic (A), Fake but Not Dangerous (FND), and Fake and Dangerous (FD). Our approach incorporates language-specific preprocessing to extract features related to sensitive keywords, sentiment polarity, and structural complexity. To enhance efficiency in large-scale detection, the Bloom filter technique has been applied for rapid parallel keyword matching. By leveraging Support Vector Machines (SVMs), Random Forest (RF), and recursive feature elimination, we have optimized classification performance across multiple languages. Experimental results on English, Italian, and Bengali datasets demonstrate that our method has achieved high detection accuracy, with the most significant improvements observed for English language content, while also yielding substantial gains in the other two languages. In addition, the BSM-DND framework is scalable in low-resource environments, adaptable to multilingual scenarios with minimal overhead, and effective in quantifying the potential societal impact of deepfake news.