Spiking Neural Networks for Log-Based Anomaly Detection: A Survey, Taxonomy, and Roadmap on Energy-Efficient Neuromorphic Log Analysis
IEEE Access · 2026 · 14 · 104519-104544
DOI: 10.1109/access.2026.3707431 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Tung Vu und 4 weiteren Autor:innen. Die Arbeit nennt 144 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Japan Science and Technology Agency (JST) Networked Exchange, United Strength for Stronger Partnerships between Japan and ASEAN (NEXUS), Japan.
Abstract (Originalsprache)
Log-based anomaly detection is essential for the reliability and security of modern software systems, yet deep learning methods face high energy costs that hinder deployment in resource-constrained settings. Spiking Neural Networks (SNNs), inspired by biological neural mechanisms, offer an energy-efficient alternative through event-driven computation and temporal spike coding. This paper presents the first survey, taxonomy, and roadmap that systematically bridges SNNs with log-based anomaly detection—two fields that have evolved largely in isolation. Because directly established SNN log-anomaly-detection literature is still nascent, we explicitly distinguish between evidence drawn directly from log-anomaly-detection studies and evidence transferred from adjacent SNN, neuromorphic, NLP, and time-series domains; claims based on transferred evidence are clearly marked throughout. We make four contributions. First, we present a comprehensive taxonomy of log-to-spike encoding schemes, SNN architectures (RSNN, CSNN, Spiking Transformer, Spiking Autoencoder), and compatible learning algorithms applicable to log anomaly detection. Second, we survey energy-efficiency evidence from the neuromorphic computing literature and provide a structured mapping between log data characteristics and SNN computational properties, establishing why SNNs are a strong candidate for log analysis. Third, we propose a joint accuracy-efficiency evaluation framework incorporating synaptic operations, energy per inference, and spike sparsity, and we provide illustrative baseline experiments via ANN-to-SNN conversion on HDFS, BGL, and Thunderbird; analytical projections (not hardware measurements) suggest 100–1, $200\times $ energy reduction potential on neuromorphic hardware at a modest 1–3% accuracy trade-off. Fourth, we identify key open challenges and future research directions, including neuromorphic hardware deployment, online continual learning, Spiking GNN architectures for log event graphs, and SNN-based explainability.
Förderung
- Japan Science and Technology Agency (JST) Networked Exchange, United Strength for Stronger Partnerships between Japan and ASEAN (NEXUS), Japan
Zitieren
APA
Vu, T., Do, D., Bui, T., Mai, D., & Pham, C. (2026). Spiking Neural Networks for Log-Based Anomaly Detection: A Survey, Taxonomy, and Roadmap on Energy-Efficient Neuromorphic Log Analysis. IEEE Access, 14, 104519-104544. https://doi.org/10.1109/access.2026.3707431
Vancouver
Vu T, Do D, Bui T, Mai D, Pham C. Spiking Neural Networks for Log-Based Anomaly Detection: A Survey, Taxonomy, and Roadmap on Energy-Efficient Neuromorphic Log Analysis. IEEE Access. 2026;14:104519-104544. doi:10.1109/access.2026.3707431
BibTeX
@article{Vu2026,
author = {Vu, Tung and Do, Duong and Bui, Thuan and Mai, Duc-Tho and Pham, Cong-Kha},
title = {Spiking Neural Networks for Log-Based Anomaly Detection: A Survey, Taxonomy, and Roadmap on Energy-Efficient Neuromorphic Log Analysis},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {104519--104544},
doi = {10.1109/access.2026.3707431}
}
EUVIMED