Learning Representations Using RNN Encoder-Decoder for Edge Security Control.

Area of Science:

• Computer Science
• Cybersecurity
• Artificial Intelligence

Background:

• Traditional whitelisting struggles with the scale of modern internet access.
• Supervised machine learning for security requires extensive labeled data, which is often impractical.
• The need for advanced, data-efficient security mechanisms is critical.

Purpose of the Study:

• To develop an unsupervised deep learning algorithm for intelligent boundary security control.
• To address the limitations of traditional whitelisting and supervised learning in network security.
• To enhance the detection of abnormal network access without relying on labeled data.

Main Methods:

• An unsupervised deep learning algorithm based on seq2seq architecture was developed.
• The model integrates recurrent neural networks and autoencoder structures.
• Data processing involved dictionary coding and sequence padding; modeling minimized reconstruction error.

Main Results:

• The proposed method achieved an Area Under the Curve (AUC) of 0.99 on a public dataset.
• Performance surpassed several classical supervised learning algorithms.
• Demonstrated efficient defense capabilities against abnormal network access.

Conclusions:

• The unsupervised deep learning approach offers an effective solution for intelligent boundary security.
• This method provides a robust defense against abnormal network traffic.
• The algorithm's superior performance highlights its potential in real-world cybersecurity applications.