Explainable AI-assisted Compliance Verification and Fraud Flagging System for 4Ps Compliance Documents Using OCR and Machine Learning Techniques

This study developed and evaluated an Explainable AI-Assisted Compliance Verification and Fraud Flagging System designed to support automated verification of 4Ps compliance documents using Optical Character Recognition (OCR), image analysis, and machine learning techniques. The increasing reliance on digital document submissions in government services such as social welfare programs has heightened the risks of document manipulation, forged records, metadata tampering, and fraudulent compliance submissions, highlighting the need for intelligent and transparent verification systems. The developed framework integrated OCR extraction using Tesseract OCR, image preprocessing, anomaly detection, explainable artificial intelligence (XAI), and ensemble-based fraud classification to improve the reliability and transparency of compliance monitoring processes. A dataset consisting of 1,200 compliance documents was generated and categorized into Valid, Needs Review, and Suspicious classes. The system supported Report Cards, Beneficiary IDs, Health Cards, and Family Development Session attendance forms. Feature engineering techniques included OCR confidence, blur score, contrast score, edge density, noise score, text length, suspicious keyword detection, and document-specific keyword matching. Linear Support Vector Classification achieved approximately 97.92% accuracy in document type classification, while Random Forest and Extra Trees classifiers achieved approximately 83–84% accuracy in fraud status classification. The framework additionally generated explainable fraud flagging reasons, anomaly indicators, audit trails, and human verification recommendations to support transparent decision-making. Findings demonstrated that the developed system can effectively support intelligent, explainable, and risk-based compliance verification in digital social welfare monitoring environments.