AN INSIDER THREAT PREDICTION AND DETECTION SYSTEM

ABSTRACT This study develops an Insider Threat Prediction and Detection System to identify and mitigate insider threats in organizational IT environments using machine learning and user behavior analytics (UBA). Leveraging the CERT r4.2 dataset, the system employs an Autoencoder for unsupervised anomaly detection, preprocessing user activity logs to extract features like access frequency and login deviations. A Streamlit-based dashboard facilitates real-time monitoring. The Autoencoder achieved an accuracy of 0.85, precision of 0.82, recall of 0.76, F1-score of 0.80, and MSE of 0.02, demonstrating robust detection with minimal false positives. Implemented locally using a virtual environment and GitHub, the system is scalable for cloud deployment. Exploratory data analysis (EDA) informed feature engineering, revealing distinct patterns in malicious activities. The study addresses challenges like data imbalance and false positives, offering a practical framework for cybersecurity. Future work includes integrating a Random Forest classifier and cloud deployment, enhancing detection capabilities.