DeepTrAnS: Unsupervised Server Telemetry Anomaly Detection
Mixer-based unsupervised anomaly detection framework for high-dimensional server telemetry, achieving 99.88% precision and 99.81% F1-score (IEEE ICSC 2025).
DeepTrAnS: Time-Series Anomaly Detection for Server Telemetry
Stack: Python, PyTorch, MLP-Mixer, Time-Series Processing, AIOps
Conference: IEEE International Intelligent Cybersecurity Conference (ICSC 2025)
DOI: 10.1109/ICSC65596.2025.11139978
Authors: Muhammad Anas Raza, Mohammad Wardat, Anyi Liu
Overview
High-dimensional server telemetry (tracking CPU load, memory paging, I/O rates, and network saturation) exhibits complex temporal correlations and sudden regime shifts.
DeepTrAnS is a pure MLP-Mixer-based unsupervised anomaly detection framework that deploys a two-phase reconstruction strategy to learn normal infrastructure operational baselines without needing labeled anomalous training samples.
Benchmark Performance
Evaluated against comprehensive production server telemetry datasets:
- Detection Accuracy: Achieved 99.88% Precision, 99.73% Recall, 99.86% AUC, and 99.81% F1-score.
- Baseline Comparisons: Drastically outperformed established baselines:
- TranAD (F1: 96.04%)
- DAGMM (F1: 93.99%)
- OmniAnomaly (F1: 90.36%)
- MERLIN (F1: 38.42%)