Hybrid DenseNet-RF Model for Automatic Crime Detection and Data-Driven Policymaking

Tax-related crimes pose significant challenges to economic stability and governance. This study presents a novel hybrid model combining DenseNet and Random Forest (RF) to analyze and classify tax-related crimes in India. By leveraging structured and unstructured datasets comprising 50,000 records from legal, financial, and governmental sources, the model extracts meaningful insights into criminal patterns. Key features include the tax amount evaded, taxpayer demographics, industry classification, and case outcomes. The model achieved an accuracy of 95.2%, with high precision (94.5%), recall (96.0%), and an F1-score of 95.2%. Compared to standalone DenseNet (93.1% accuracy) and RF (91.8% accuracy), the hybrid model demonstrated superior performance. Feature importance analysis revealed “Tax Amount Evaded” and “Frequency of Offenses” as the most critical attributes influencing predictions. Insights derived from the model guided the formulation of actionable legal strategies, including enhanced penalties for repeat offenders and improved monitoring of high-risk industries. The findings underscore the potential of hybrid deep learning and ensemble methods in automating tax crime detection and shaping datadriven policymaking.