Energy-Aware Federated Learning for IoT Intrusion Detection Using Latent Feature Encoding
IEEE Access · 2026 · 14 · 71505-71524
DOI: 10.1109/access.2026.3691274 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Vipin Kumar Kushwaha und 3 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 54 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
With the rapid rise of the Internet of Things (IoT), the attack surface of today’s cyber-physical infrastructure has expanded dramatically. Therefore, there is an urgent need for scalable, privacy-preserving solutions for intrusion detection systems. Recently, a new paradigm, Federated Learning (FL), has emerged as a means for providing distributed IoT security. Despite this, real-world deployments face considerable hurdles due to heterogeneous device capabilities, non-independent and identically distributed data distributions, and significant communication overheads. To help mitigate these challenges, we present our energy-aware Federated Intrusion Detection Framework, which combines latent feature encoding with three FL approaches. A compact encoder converts multi-source IoT traffic into a 64-dimensional latent vector, thus enabling efficient on-device training across a broad range of diverse data types. To further illustrate the effectiveness of latent feature encoding in improving the performance of FL models, we have evaluated our baseline centralized model and three FL models (FedAvg, CFL, and FedProx). All methods performed well under heterogeneous conditions. However, CFL and FedProx both exhibited significantly greater performance than FedAvg, with accuracies of 0.967455 and 0.967914, respectively, compared to the accuracy of FedAvg (0.840155). To evaluate the feasibility of deploying our framework on resource-constrained IoT devices, we have developed an energy-aware simulation that models battery capacity, compute load, communication costs, and client failures. For the combined dataset, the energy consumed by the full FL pipeline was 1055.3694 units, with a total communication volume of 199,106,400 bytes exchanged during five rounds. Ultimately, this analysis demonstrates the efficacy of latent feature encoding in improving the performance of FL models and provides valuable insights into energy consumption patterns for the deployment of real-world IoT intrusion detection systems.
Zitieren
APA
Kushwaha, V. K., Verma, D. K., Yadav, S. P., & Gupta, H. (2026). Energy-Aware Federated Learning for IoT Intrusion Detection Using Latent Feature Encoding. IEEE Access, 14, 71505-71524. https://doi.org/10.1109/access.2026.3691274
Vancouver
Kushwaha VK, Verma DK, Yadav SP, Gupta H. Energy-Aware Federated Learning for IoT Intrusion Detection Using Latent Feature Encoding. IEEE Access. 2026;14:71505-71524. doi:10.1109/access.2026.3691274
BibTeX
@article{Kushwaha2026,
author = {Kushwaha, Vipin Kumar and Verma, Deepak Kumar and Yadav, Satya Prakash and Gupta, Himanshu},
title = {Energy-Aware Federated Learning for IoT Intrusion Detection Using Latent Feature Encoding},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {71505--71524},
doi = {10.1109/access.2026.3691274}
}
EUVIMED