Detecção de ataques de injeção de dados falsos em sistemas de potência utilizando autoencoders com mecanismo de atenção.

Abstract

This dissertation proposes a method for detecting False Data Injection (FDI) attacks in power systems using autoencoders with a multi-head attention mechanism. FDI attacks pose a significant threat to the security and stability of modern electrical grids by manipulating measurement data to deceive control systems undetected. The developed model combines the ability of autoencoders to learn normal operation patterns with the efficiency of the multi-head attention mechanism to identify subtle anomalies caused by attacks. The methodology was validated on systems of varying complexity (IEEE-14, 30, 118, and 300 buses) using data generated by MATPOWER. The results demonstrated high detection efficacy, with accuracy exceeding 97% for attack variations of ±50% and ±25%, and robust performance even for more subtle attacks (±15%). The proposed approach outperforms traditional methods by eliminating the need for labeled data and complex system models, standing out as a scalable and adaptable solution for critical infrastructure security.