MaskNet: Enhancing Crime Event Detection with Feature Masking and Dynamic Attention

Crime event detection is critical in modern security systems, requiring precise identification and prediction of criminal behavior from large data sets to provide effective law enforcement and public safety. This paper offers MaskNet, a novel technique that improves criminal event detection by combining dynamic multi-head attention and feature masking in Transformer models. Traditional attention systems frequently fail to prioritize key aspects, which reduces detection efficiency. MaskNet overcomes this restriction by utilizing dynamic head selection, which allows the model to adaptively focus on the most informative attention heads. Furthermore, feature masking highlights crucial traits while suppressing noise, resulting in improved overall performance. We additionally provide the Bengali Crime Event Detection Dataset, which includes 9,999 labeled crime event examples chosen from online news portals and annotated by professionals. This dataset includes useful information about crime incidents in Bangladesh, notably in Bengali, and contributes to regional crime detection research. MaskNet outperforms the baseline Transformer models with 97.69% accuracy, making it suitable for real-world crime analytics and predictive methods. Code and Dataset Available at https://github.com/MIthun667/Crime-Event-Detection