XAInomaly: Explainable and Interpretable Deep Contractive Autoencoder for O-RAN Traffic Anomaly Detection

Explainability is critical for AI in network management: The paper demonstrates that traditional deep learning models for detecting network problems are "black boxes," making it difficult for network operators to understand why the AI flagged an issue. The XAInomaly framework addresses this by using a technique called "fastshap-C" to explain its decisions, which is essential for operators to trust and act on AI recommendations.

Semi-supervised learning reduces data labeling burden: The Deep Contractive Autoencoder uses semi-supervised learning, meaning it can learn from mostly unlabeled network data rather than requiring extensive manual labeling of normal versus abnormal traffic. This is practical for real-world deployment since labeling all network anomalies is time-consuming and expensive.

Interpretability enables effective network management: The paper emphasizes that for complex, disaggregated networks like O-RAN (which combine equipment from multiple vendors), operators need to not only detect problems accurately but also understand what caused them. This interpretability is positioned as essential for effective decision-making, not just a nice-to-have feature.

Generative AI can model normal behavior to spot deviations: Rather than trying to learn what anomalies look like, the framework learns what normal network traffic looks like, then flags significant deviations. This approach is more practical since normal patterns are more consistent and easier to model than the infinite variety of possible problems.