Transformer-based anomaly detection in internet of things networks: a systematic review, meta-analysis, and proposed tiad-net architecture (2017-2025)

Abstract

The rapid growth of Internet of Things (IoT) devices in areas such as healthcare, smart cities, transportation, and industrial automation has generated massive amounts of multivariate time-series data. Detecting anomalies in this data is essential for identifying cyber-attacks, device failures, and sensor malfunctions. Recently, deep learning techniques, especially Transformer-based models, have shown significant improvements in anomaly detection performance due to their ability to capture long-range temporal dependencies in complex datasets. This study presents a systematic review and meta-analysis of deep learning-based anomaly detection methods for IoT time-series data published between 2017 and 2025 using a PRISMA-compliant methodology. A total of 94 research papers were selected for qualitative analysis, while 71 studies were included in the quantitative meta-analysis. The risk of bias was evaluated using the Cochrane framework. The findings indicate that Transformer-based approaches outperform traditional LSTM Autoencoder methods with an average improvement of 6.3 F1-score points. Additionally, the proposed TiAD-Net model, which combines sparse self-attention with temporal convolution blocks, achieved high detection accuracy across benchmark datasets. The study also highlights major challenges including computational cost, limited labeled datasets, and edge-device deployment constraints, while identifying future research directions for IoT anomaly detection systems.