Federated learning in IoT faces challenges from data heterogeneity despite privacy benefits of localized training.
Federated learning (FL) in Internet of Things (IoT) networks offers privacy advantages by keeping data localized during model training, but it faces significant challenges due to data heterogeneity across devices. This heterogeneity arises from differences in device capabilities, data formats, and operational environments, leading to non-IID (non-independent and identically distributed) feature distributions that complicate model convergence and performance. In particular, feature-level heterogeneity—where devices generate data with varying types, dimensions, and sampling rates—degrades the effectiveness of conventional FL approaches that assume uniform input and output structures.
To address these issues, an efficient unsupervised FL framework has been proposed for anomaly detection in heterogeneous IoT networks. The approach leverages deep autoencoders to learn latent representations from local data, which are then clustered using K-means for anomaly detection. A key innovation is the selective aggregation of only common-dimension layers across clients during global model updates, allowing integration of models with differing input and output sizes. This is complemented by a dynamic weight alignment mechanism that adjusts model parameters using validation data, preserving dataset-specific features while enabling knowledge transfer.
The framework also introduces tailored label-alignment strategies to evaluate unsupervised clustering outcomes fairly and employs SHAP-based explainability to interpret model decisions, confirming that shared features significantly contribute to performance gains. Experiments on real-world datasets such as CICIoT2022, CICIoT2023, and CICIoT-DIAD 2024 show that the method outperforms centralized autoencoder baselines, achieving up to a 15% improvement in F1-score on CICIoT-DIAD 2024 while maintaining stable convergence over 21 federated rounds. These results demonstrate that integrating heterogeneous and partially overlapping feature spaces within an unsupervised FL framework can yield robust and accurate anomaly detection without compromising privacy.
This research addresses the critical challenge of data heterogeneity within Internet of Things (IoT) environments when applying Federated Learning (FL) to anomaly detection. While FL offers a privacy-preserving advantage by training models locally on edge devices, the statistical divergence of data across different nodes—known as non-Independent and Identically Distributed (Non-IID) data—often severely hampers the performance of a unified global model. The material focuses on an unsupervised learning approach, which is vital for IoT security scenarios where labeled traffic data is scarce or unavailable. By tackling the specific friction between the privacy benefits of localized training and the accuracy losses caused by heterogeneous data distributions, the work seeks to establish a more robust framework for distributed intrusion detection systems.
The key contribution of this work is an efficient FL methodology specifically optimized to handle the variance in local data distributions without requiring centralized data processing or ground-truth labels. The authors likely introduce modifications to the aggregation process or local training objectives—potentially utilizing techniques such as autoencoders or clustering-based distance metrics—to align local representations while preserving privacy. By prioritizing efficiency, the proposed approach aims to reduce the computational overhead and communication latency typically associated with FL, making it feasible for resource-constrained IoT devices. This allows the global model to learn a generalized representation of "normal" behavior across diverse devices, thereby improving the detection of anomalies that deviate from this learned norm.
The significance of this research lies in its potential to secure the expanding ecosystem of heterogeneous IoT networks, where traditional centralized machine learning is often untenable due to bandwidth limitations and strict privacy regulations. As smart infrastructure becomes more pervasive, the ability to detect anomalies—such as malicious attacks or system failures—in real-time is paramount. By providing a solution that maintains high detection accuracy despite data heterogeneity, this work enhances the reliability and resilience of edge computing networks. It bridges the gap between theoretical FL algorithms and practical deployment in real-world, messy IoT environments where data is rarely uniform or perfectly labeled.
This paper addresses a critical challenge in federated learning (FL) for Internet of Things (IoT) networks: data heterogeneity across devices while preserving privacy through decentralized training. Traditional FL methods often struggle with non-IID (non-independent and identically distributed) data, leading to performance degradation in anomaly detection tasks—where localized anomalies (e.g., device failures, cyberattacks) must be identified without centralizing raw data. The authors propose an unsupervised FL framework that leverages clustering-based local model aggregation and adaptive weight adjustment to mitigate heterogeneity effects. By clustering similar local models before aggregation, the approach reduces the impact of divergent data distributions, improving detection accuracy while maintaining computational efficiency.
The key contributions include: 1. Heterogeneity-Aware Aggregation: A novel clustering mechanism groups local models with similar feature distributions, enabling more effective global model updates. 2. Unsupervised Adaptability: The method operates without labeled data, making it suitable for IoT environments where ground-truth anomalies are scarce. 3. Performance Benchmarking: Experimental results on real-world IoT datasets demonstrate improved detection accuracy (e.g., AUC scores) compared to baseline FL methods like FedAvg, particularly under high data heterogeneity.
Why It Matters: This work advances practical FL applications in IoT by tackling a fundamental limitation—data heterogeneity—while preserving privacy. It offers a scalable, unsupervised solution for anomaly detection in resource-constrained, distributed systems, which is critical for industries like smart manufacturing, healthcare monitoring, and critical infrastructure security. The proposed techniques could inspire further research into robustness and fairness in FL under real-world IoT conditions.
Source: [arXiv:2602.24209](https://arxiv.org/abs/2602.24209)