Wireless Sensor Networks (WSNs) provide reliable connectivity and data-gathering capabilities, but accurate node localization remains a persistent challenge, especially in 3D environments where altitude variations complicate distance estimation. Most existing 3D localization methods rely on a minimum of four anchor nodes (ANs), which significantly increases deployment costs. To reduce this cost, a single-anchor 3D localization approach is proposed that uses a virtual-anchor projection strategy combined with a Lévy Flight-Enhanced Adaptive Butterfly Optimization Algorithm (LF-EABOA) to improve accuracy and efficiency. As soon as a mobile target node falls within the communications range of a single anchor node, six virtual anchor nodes at the same distance are created around the anchor node. After that, an adaptive selection mechanism will select the best virtual anchors that will provide robust localization without the need for deployment of physical anchors. This method explicitly models noise and interference using a log-distance path-loss model, including shadowing effects and a degree of irregularity (DOI) and non-line-of-sight (NLOS) penalties. In order to verify the proposed method, extensive simulation experiments were conducted using an average of 80 nodes to represent medium-scale configurations, which demonstrated that the proposed LF-EABOA method produced a mean localization error of 0.1895 m, which was significantly less than existing traditional methods and was also associated with substantially lower costs. Statistical analysis of confidence intervals and mobility sensitivity demonstrate that the proposed method is robust.
The increasing adoption of smart home technologies has intensified the need for intelligent and privacy-preserving Human Activity Recognition (HAR) systems. Centralized learning approaches expose sensitive behavioral data to privacy risks, while existing federated learning methods remain vulnerable to inference attacks and face challenges arising from heterogeneous and resource-constrained home environments. To address these limitations, we propose FedPrivHAR-LLM, a unified framework that integrates federated learning, differential privacy, homomorphic encryption, and Large Language Models (LLMs). Federated learning enables distributed model training without raw data sharing, differential privacy provides formal statistical guarantees against re-identification, and homomorphic encryption secures model aggregation in the encrypted domain. Beyond secure training, LLM-based semantic reasoning is incorporated to enhance context-aware interpretation and disambiguation of human activities. Experimental evaluations on three benchmark HAR datasets adapted to smart-home settings (CASAS, Opportunity, and PAMAP2) show that FedPrivHAR-LLM achieves up to 90.2
The increasing demand for task-specific adaptation of large language models (LLMs) on private, geographically distributed data has led to a growing interest in federated fine-tuning of LLMs (FedLLM). However, existing FedLLM methods often face immense challenges, such as non-IID data distributions, device heterogeneity, and high communication costs, with mathematically inconsistent LoRA adapter fusion due to naive averaging under varying ranks. This paper proposes a lightweight two-level hierarchical framework for federated LoRA fine-tuning. The proposed methods group client devices based on the cosine similarity of recent model updates. Within each cluster, a cluster head merges and computes the true weighted average of the LoRA updates using stacking. Subsequently, inter-cluster aggregation is performed at the central server, which uses weighted fusion followed by per-layer truncated SVD, producing a compact global adapter. The performance of the proposed method is evaluated using the RoBERTa base and the DeBERTa-v3 base, applying to the SST-2 and QNLI datasets, respectively. The results suggest that the proposed method achieves lower network traffic to target accuracy. For SST-2 dataset, the deduction in network traffic is 5.9
The Internet of Things has increased its application areas, and most of the applications are battery-powered IoT devices. The energy optimization has been the core issue in such miniature devices. Although there has been a lot of research going on in this direction. Researchers have provided many solutions at the device level and network level. The solutions include design aspects in hardware and at the protocol level. We expect 125 billion IoT devices by 2030. Therefore, energy-efficient approaches are required at the hardware and software level. We expect more emphasis has to be given to green computing in these devices; this will help us to reduce the environmental impact and also help us to reduce the energy consumption in all such networks. This paper discusses the energy optimization techniques and utilization of non-conventional energy resources, and in particular the green internet of things (Green IoT), by having optimized energy management techniques for IoT environments. This review comparatively examines energy-harvesting and energy-saving approaches across device and network levels. It highlights their operating principles, practical constraints, and suitability for Green IoT applications.
Accurate identification of Internet of Things (IoT) device types supports security monitoring and asset management. Fixed observation windows can fragment communication cycles, while supervised methods may require retraining across deployment environments. We propose PerSiFi (Periodic Similarity-based Fingerprinting), a method for continuous device identification from idle-state traffic without supervised classifier training. PerSiFi extracts candidate flow periods in the frequency domain and validates them through autocorrelation and stability screening. It then aligns cycle segments and separately fuses payload similarity digests and phase–direction–length structures, retaining period information in flow-level sub-fingerprints. During identification, maximum-weight one-to-one matching compares sub-fingerprints from nested traffic prefixes with an enrolled library. Score and margin thresholds determine whether to output an identity, continue observation, or abstain. On Mon(IoT)r, PerSiFi achieves 97.0
