STS-AT: A Structured Tensor Flow Adversarial Training Framework for Robust Intrusion Detection.

Area of Science:

• Computer Science
• Cybersecurity
• Artificial Intelligence

Background:

• Network intrusion detection is crucial for cybersecurity.
• Current methods suffer from manual feature engineering and vulnerability to adversarial attacks.
• Deep learning models often lose discriminative information and are susceptible to sophisticated threats.

Purpose of the Study:

• To propose STS-AT, a novel network intrusion detection method.
• To address the limitations of manual feature engineering and adversarial vulnerabilities in current systems.
• To enhance the accuracy, robustness, and efficiency of network intrusion detection.

Main Methods:

• Structured tensor encoding to convert raw traffic into numerical representations.
• A hierarchical deep learning model combining CNN and LSTM for spatial-temporal feature learning.
• Multi-strategy adversarial training to enhance model robustness against attacks.

Main Results:

• Achieved 99.6% accuracy in normal traffic classification on the CICIDS2017 dataset.
• Significantly outperformed Random Forest (93.1%) and Support Vector Machine (84.7%).
• Defense accuracy against adversarial attacks increased to over 96.8%, compared to 24.4% for undefended models, with a 67.6% reduction in training time.

Conclusions:

• Structured tensor encoding effectively preserves original traffic information.
• The hierarchical model enables comprehensive feature learning.
• Multi-strategy adversarial training improves efficiency and ensures robust defense against cyber threats.