A Privacy-Preserved Variational-Autoencoder for DGA Identification in the Education Industry and Distance Learning

Area of Science:

• Cybersecurity
• Machine Learning
• Data Privacy

Background:

• Domain generation algorithms (DGAs) are used by malware to evade detection by frequently changing command and control (C&C) server domains.
• Training machine learning models to combat DGAs requires sharing sensitive data, raising privacy and integrity concerns.
• The rise of remote learning, accelerated by the pandemic, has increased the attack surface for DGA-based threats.

Purpose of the Study:

• To propose a novel privacy-preserving machine learning approach for DGA detection.
• To address data integrity and privacy concerns associated with training DGA detection models.
• To apply the proposed method to the education industry, focusing on distance learning environments.

Main Methods:

• Development of a privacy-preserved variational autoencoder tailored for DGA detection.
• Integration of secured multi-party computation (SMPC) to enable machine learning on encrypted data.
• Utilizing a Siamese variational autoencoder algorithm for enhanced feature extraction from encrypted metadata.

Main Results:

• The proposed system successfully applies machine learning techniques to encrypted data and metadata.
• The method demonstrates improved training stability and high generalization performance in DGA categorization.
• Achieved remarkable categorization accuracy in identifying DGA-based threats within the education sector.

Conclusions:

• The novel privacy-preserving variational autoencoder combined with SMPC offers a secure solution for DGA detection.
• This approach effectively mitigates privacy and data integrity risks in machine learning model training.
• The system provides a robust and accurate method for enhancing cybersecurity in remote learning environments.