Executive Summary
This thesis evaluates the efficacy of Long-Short Term Memory (LSTM) autoencoders for detecting traffic anomalies in 5G wireless networks supporting energy communications systems at Marine Corps Air Station (MCAS) Miramar. The research compares LSTM performance against a simpler Support Vector Machine (SVM) algorithm using a simulated 5G dataset with specific attacks. Findings indicate that the LSTM model, despite its complexity, did not outperform the SVM in anomaly detection for the tested data representations.
Why It Matters
This document is crucial for defense analysts as it addresses cybersecurity vulnerabilities in 5G networks supporting critical energy infrastructure on military bases, directly impacting operational resilience and national security.
Key Takeaways
- MCAS Miramar is implementing 5G wireless networks for its Energy Management System (EMS) to enhance resilience and manage distributed energy resources.
Source evidence · PDF page 8
The Energy Management System (EMS) in place at Marine Corps Air Station(MCAS) Miramar is undergoing a restoration process. The aim is to establish wirelesscommunication between the EMS facility and the backup generators and photovoltaicarrays that are distributed across the station using the Verizon 4G LTE/5GUltra-Wideband network.
- The LSTM autoencoder, while capable of pattern analysis, showed poor performance in detecting 5G-specific cyber anomalies compared to a simpler SVM model.
Source evidence · PDF page 59
It is apparent that, in the case ofthese representations of the data, the SVM performs better with a higher accuracy and recallin each category. Furthermore, when trained on the same hardware, the LSTM takes severalorders of magnitude longer to train than the SVM.
- Pre-processing techniques significantly impacted the LSTM's effectiveness, with the model performing best when packet size was included in features, but overall, SVM yielded superior accuracy and recall.
Source evidence · PDF page 62
Overall, the LSTM did not perform well, likely due to the pre-processing techniques chosen.While experimenting with other input methods, such as raw packet input, might performbetter, the SVM, a much simpler algorithm, performed better and further gains will be slim.
Strategic Relevance
Securing 5G networks for military energy systems is vital for maintaining operational continuity and resilience against cyber threats. This research highlights challenges in applying advanced machine learning for anomaly detection in such critical infrastructure, informing future cybersecurity strategies and technology adoption for defense assets.