AI-Based Detection of Phishing Attacks in Digital Payment Systems Using Explainable Machine Learning

Abstract

India's rapid adoption of Unified Payments Interface (UPI), mobile banking, digital wallets, and other digital payment channels has expanded both financial inclusion and the attack surface available to cybercriminals. AI-enabled phishing, voice cloning, facial manipulation, and identity impersonation increasingly challenge conventional rule- and signature-based security mechanisms. This paper proposes an integrated, explainable machine-learning framework for detecting phishing and deepfake-related fraud in digital payment environments. The framework combines transaction-behavior analysis, phishing indicators, multimodal deepfake evidence, and an Explainable AI (XAI) layer based on SHAP and LIME. The data-processing pipeline includes multi-source collection, schema harmonization, validation, duplicate removal, leakage-aware preprocessing, feature engineering, stratified splitting, model training, hyperparameter optimization, and cross-dataset evaluation. XGBoost is positioned as a primary tabular fraud-detection model, while CNN/LSTM/transformer-based components are proposed for visual, temporal, audio, and phishing-content analysis. Because the supplied manuscript does not report completed experimental measurements, numerical performance figures are treated as targets rather than observed results. The proposed framework is intended to improve detection effectiveness, interpretability and operational trust while providing a scalable basis for future validation in Indian digital-payment ecosystems.