LMP-GAN: Out-of-Distribution Detection for Non-Control Data Malware Attacks.

Area of Science:

• Machine Learning
• Cybersecurity
• Statistical Inference

Background:

• Anomaly detection, particularly out-of-distribution (OOD) detection, is a critical machine learning task.
• OOD detection is semi-supervised, training only on inlier samples without modeling outlier distributions.
• Cyber-physical systems face novel threats like non-control data (NCD) attacks, evading traditional malware detection.

Purpose of the Study:

• To develop a novel Generative Adversarial Network (GAN)-based OOD detection network.
• To protect cyber-physical signal systems from sophisticated NCD Trojan malware attacks.
• To leverage principles from statistical inference, specifically the locally most powerful (LMP) test.

Main Methods:

• Designed a novel GAN-based OOD detection network, termed LMP-GAN.
• Trained the discriminator to generate OOD samples that maximize inlier alteration while evading detection.
• Inspired by the classical locally most powerful (LMP) test for statistical inference.

Main Results:

• The proposed LMP-GAN demonstrates effective OOD detection capabilities.
• Experimental results show superior performance compared to state-of-the-art anomaly detection methods.
• The network successfully identifies novel NCD attacks that bypass conventional detection techniques.

Conclusions:

• LMP-GAN is a highly effective OOD detector for cyber-physical systems.
• The method provides robust protection against advanced NCD malware.
• The approach validates the benefits of integrating LMP principles into GAN-based anomaly detection.