Synthetic Communication Anomaly Dataset for Prototype-Guided Contrastive Detection

This dataset contains 5,000 synthetically generated communication sessions designed to simulate normal and anomalous traffic in distributed AI systems. Each sample consists of 20 continuous-valued statistical features commonly observed in low-resource communication environments, such as jitter, packet rate deviation, and entropy variation. Normal data accounts for 85% of the samples and is drawn from clustered Gaussian distributions, while the remaining 15% represents anomalies generated from a broader uniform distribution to emulate unpredictable faults and adversarial disruptions. The dataset supports research in anomaly detection, particularly in scenarios with limited supervision, and was used to validate a prototype-guided contrastive learning framework for one-class classification. All data is normalized and split into training and test sets with provided ground truth labels for reproducibility.