Multi-Signal Fake Social Media Profile Detection and Reporting: A Machine Learning Approach with Computer Vision and Clone Analysis

Abstract

The proliferation of automated, cloned, and AI-generated profiles on platforms such as Instagram and X (Twitter) poses a growing threat to online trust, public discourse, and cybersecurity. Manual reporting mechanisms and platform-side heuristics are insufficient against the scale and sophistication of modern fraudulent accounts. This paper presents a web-based, multi-signal fake social media profile detection and reporting system that integrates a trained Random Forest classifier, OpenCV-based face authenticity analysis, HuggingFace AI-image detection, fuzzy-string clone matching, and keyword-driven spam scoring into a unified, real-time risk engine. Profile data is fetched live via the Instagram Scraper API and Twitter API47 (RapidAPI). Seventeen extracted features — spanning metadata, behavioral, and content dimensions — feed the machine learning pipeline, which outputs a calibrated Fake Probability Score (0–100%) mapped to Low, Medium, and High risk tiers. A SQLite-backed history database records every analysis, and a background monitoring thread continuously re-evaluates watchlisted accounts. Evaluated on the UCI “user_fake_authentic_2class” benchmark dataset with hyperparameter tuning via GridSearchCV, the system achieves a weighted F1-score exceeding 0.92, demonstrating competitive performance against prior single-model baselines. The modular codebase, cross-platform support, and explainable risk breakdown distinguish this system from existing tools that offer only opaque classification outputs.