This paper presents an approach to anomaly detection in network performance data collected through the perfSONAR-based Performance Measurement Platform (PMP), in GN5-2 Work Package 6, Task 3. Using unsupervised machine learning techniques, specifically autoencoders, the study focuses on identifying deviations in latency distribution without relying on labelled datasets. Two approaches are evaluated: direct histogram-based learning and statistical feature extraction. Results demonstrate that the first proposed method successfully detects both transient anomalies and sustained distribution shifts, highlighting its potential for real-time network monitoring.