Intent2Vec: Contrastive Learning for Predictive and Interpretable Intent Violation Detection in SDN

Intent-Based Networking enhances Software-Defined Networks by enabling high-level policy automation, yet existing monitoring systems remain reactive and lack semantic awareness of declared intents. This paper presents Intent2Vec, a contrastive learning framework for predictive and interpretable intent-violation detection. The model jointly embeds flows and intents into a shared latent space, where proximity reflects compliance and divergence signals violations. Evaluated on the SDNFlow dataset, Intent2Vec achieves 98.7% accuracy, 99.5% precision, and 98.4% recall, confirming its predictive reliability. Embedding visualizations (t-SNE, UMAP) show clear separation between compliant and violating flows, and intent-level metrics, precision 0.9961 and safety score 0.9961 demonstrate strong semantic alignment. Overall, Intent2Vec bridges the gap between declared intents and observed behavior, offering a proactive, explainable, and controller-ready foundation for intent-driven SDN security.