ARCHIVES

The proliferation of fake and bot-driven profiles across social media platforms poses significant threats to digital trust, user safety, and information integrity. This paper presents a comprehensive fake profile detection system built as a Flask- based web application that employs multi-layer ensemble ma- chine learning to classify profiles across ten major social media platforms: Instagram, Facebook, X (Twitter), LinkedIn, GitHub, Discord, YouTube, TikTok, Reddit, and Snapchat. The system implements a two-stage tiered prediction architecture wherein a quick scan using ensemble classifiers (Random Forest, XGBoost, SVM, and HistGradientBoosting) provides immediate results for high-confidence cases, while borderline predictions trigger a deep analysis pipeline incorporating live API enrichment, avatar photo forensics, username pattern analysis, bio natural language processing, behavioral fingerprinting, and cross-platform identity verification. The ensemble employs stacking with LightGBM as a meta-learner, isotonic regression calibration, Optuna hy- perparameter tuning, and SMOTE/ADASYN class balancing. Trained on a combined dataset of 7,400 Instagram profiles (5,000 real-world and 2,400 synthetic), the system achieves 98.4% classification accuracy. The application features SHAP-based explainability, Population Stability Index (PSI) drift monitoring, user feedback loops for continuous model improvement, and com- prehensive REST API endpoints with Swagger documentation. Experimental results demonstrate the effectiveness of the multi- layer forensic approach in improving prediction confidence for both obvious and borderline profiles.