Federated learning (FL) offers significant advantages in preserving data privacy and enhancing communication efficiency, making it especially suitable for Internet of Things (IoT) networks supported by unmanned aerial vehicles (UAVs). However, most existing FL approaches rely on assumptions of uniformly distributed, well-labeled, and large-scale datasets-conditions that rarely hold in practical UAV-based IoT scenarios. These environments typically feature small-scale, non-independent and identically distributed (non-IID), and dynamically changing data. To address these challenges, we propose a novel self-supervised federated unsupervised learning (FUL) framework tailored for UAV-assisted IoT systems. The proposed framework comprises three key components: (1) a realistic UAV data collection model that considers limited onboard storage and mobility constraints; (2) a robust local training strategy that incorporates self-supervised regularization and a centered kernel alignment (CKA)-based similarity loss to mitigate the effects of data heterogeneity and rapid distribution shifts; and (3) an importance-aware hybrid normalized aggregation method at the global server, which leverages model divergence-based metrics to evaluate local model reliability and integrates both current and historical gradient information for stable model updates. Experimental results demonstrate that our framework achieves classification accuracies of 30.5%, 62.8%, and 70.5% under memory constraints of 500, 1000, and 2000 samples, respectively—outperforming the best baseline by 18.1%, 26.5%, and 9.1% under the same conditions. These results highlight the effectiveness of the proposed FUL framework in handling data heterogeneity and dynamic sample variations inherent in realistic UAV-enabled IoT applications.

This paper proposes a new design-technology cooptimization framework that expedites circuit optimization by utilizing the neural compact modeling (NCM) and a data-driven SPICE simulation. An efficient retargeting strategy of NCM and its improved design capability through a direct data driven SPICE simulation were leveraged at the industry level in response to increasingly challenging current development situations. To facilitate rapid feedback for extensive trial and error in technology optimization, the NCM swiftly fine-tune itself using pre-trained model. Then, the data interpolation and derating techniques are utilized to provide the same design environment as before such as instance binning, process variations, and layout dependent effects. Demonstrating the robustness of our framework, we achieved a 95% reduction in PDK release time while maintaining model consistency and performance at a mid-scale design of $\mathbf{1 5 k}$ transistors, with no SPICE run time and accuracy loss. This solution allows for rapid incorporation of process changes into the design, supporting quick path-finding during a design-technology co-development.

Federated Learning (FL) enables collaborative training across distributed edge devices. However, it struggles with statistical heterogeneity in non-IID scenarios, which degrades overall model performance. Clustered FL addresses this issue by grouping clients with similar data distributions, allowing clients within each cluster to the similar data for more personalized training. However, existing clustered FL methods face challenges in accurately identifying data similarities and often introduce significant communication overhead and computational costs. In this paper, we propose a clustered FL framework (ProCFL) that exploits local prototypes during the initialization phase for one-shot identification of client data similarities, guiding subsequent federation. Each client computes a local prototype representing its data distribution using a common model and uploads it to the server. The server employs Singular Value Decomposition (SVD) to extract principal vectors from these prototypes, addressing label misalignment and enabling pairwise angular similarity computation. To achieve better cluster assignments, ProCFL incorporates a hierarchical soft clustering mechanism that forms overlapping cluster sets, promoting knowledge sharing across clusters. We evaluate our method under Label Skew and Feature Skew non-IID scenarios using multiple datasets, including Fashion-MNIST, CIFAR-10, and Digit-5. The results show that ProCFL achieves higher test accuracy than existing methods while reducing communication overhead from 11.71 MB to 0.27 MB and clustering time from 6.07 s to 0.44 s.

Activity recognition is essential in computer vision applications, such as smart homes and healthcare services. While RGB images have been widely used in this area, they pose challenges related to privacy invasion and environmental constraints. To address these issues, some research has explored using 3-D light detection and ranging (3-D LiDAR) to collect 3-D point cloud data for activity recognition. However, the high-computational cost and large model parameters required for processing 3-D point clouds remain major limitations. To overcome these challenges, we propose a novel multiclass activity recognition system based on skeleton extraction from depth images collected by 3-D LiDAR. First, we use 3-D LiDAR to collect depth images of ten distinct activities, such as walking, falling, and squatting. Next, we process these depth images using our proposed multiple dynamic nonlinear mapping (MDNLM) method. The MDNLM method enhances the clarity of human body details by adjusting the color distribution of depth values based on the human position, ensuring that more colors are allocated to specific regions of the human body. This enhancement allows a fine-tuned algorithm to extract skeleton joints accurately from the mapped images. Finally, the extracted skeletons are fed into a convolutional neural network combined with a long short-term memory network (CNN+LSTM) for multiclass activity recognition. Our proposed method achieved 100.0% accuracy for a 2-class classification task (fall detection), 99.0% accuracy for a 7-class classification task, and 94.7% accuracy for a 10-class classification task.

Detecting small floating targets on the sea surface has long been a major challenge in radar signal processing. Recently, deep learning (DL) has attracted considerable attention for its potential to improve detection probability. However, its performance heavily relies on the availability of sufficiently labeled datasets, which are often difficult to acquire in complex sea clutter environments. Therefore, this letter introduces the Complex-Valued Hybrid Spatio-Temporal Network (CVHSTNet), a novel maritime radar target detection method designed for low-data scenarios that utilizes time-frequency (TF) representations of radar echoes as inputs. To mitigate the overfitting issue, CVHSTNet is intentionally designed with a shallow architecture, integrating a three-layer complex-valued convolutional neural network (CV-CNN) with a one-layer complex-valued bidirectional long short-term memory network (CV-BiLSTM). Unlike existing real-valued models that overlook phase information, our method operates directly on complex-valued data to capture the complete signal representation. More importantly, this hybrid architecture enables the network to effectively exploit both spatial and temporal characteristics, thereby further enhancing feature representations. Comprehensive experiments on 40 datasets from the IPIX database demonstrate that, with only 50 samples per range cell for training, the proposed method achieves a detection probability exceeding 90% in 37 out of 40 datasets, with a false alarm rate (FAR) of 10−3. To the best of our knowledge, this is the first time a DL-based approach has demonstrated the ability to distinguish between small floating targets and sea clutter under limited labeled radar data conditions.

High Altitude Platform Stations (HAPS) are instrumental in wireless communications, providing enhanced connectivity and extensive coverage by complementing ground-based infrastructure where its expansion is limited. HAPS enhance wireless communications by employing advanced beamforming technology, traditionally based on 2D with free space path loss models. Addressing the inadequacy of these models in different areas, our research introduces a novel beamforming strategy that incorporates detailed 3D geographic and architectural information. This approach models line-of-sight (LOS) and non-line-of-sight (NLOS) conditions with 3D information and optimizes beamforming patterns using Deep Reinforcement Learning (DRL). Our results indicate that by incorporating 3D information, the beamforming performance in terms of average throughput and SINR is markedly enhanced across all user equipments (UEs), compared to traditional 2D approaches. By taking 3D information into account, beamforming in HAPS systems becomes more equitable.

An optical fiber distributed acoustic sensing (DAS) system for large infrastructure vibration monitoring is proposed in this work. To meet the requirements of measurement range, spatial resolution, and real-time performance of the monitoring network, the acrlong RE algorithm is proposed to optimize the recovery of large signals for the DAS monitoring of large-scale infrastructure structure monitoring networks. Furthermore, the technology is applied to heavy rail track defect detection, where existing track-side communication cables are used to directly monitored vibration signals with the DAS system. Multiple characteristic parameters are combined to form a multidimensional eigenvector, and then combined with the acrlong ML algorithm to enable the recognition of typical track defects along the heavy-haul railway. The experimental results demonstrate that the recognition and classification of typical track defects, such as acrlong RCF, corrugation, and unsupported sleepers. The real-time detection of track defects in this work can be used as a crucial basis for workers to maintain and repair the railway. Finally, a long-term real-time online monitoring method is proposed in this work for vibration monitoring of large-scale infrastructures with large-amplitude/low-SNR signals using existing track-side communication cables, without any additional sensor arrangement.

Federated Learning is now widely used to train neural networks under distributed datasets. One of the main challenges in Federated Learning is to address network training under local data heterogeneity. Existing work proposes that taking similarity into account as an influence factor in federated learning can improve the speed of model aggregation. We propose a novel approach that introduces Centered Kernel Alignment (CKA) into loss function to compute the similarity of feature maps in the output layer. Compared to existing methods, our method enables fast model aggregation and improves global model accuracy in non-IID scenario by using Resnet50.

Convolutional neural network (CNN) is one of the most important tools to accomplish high-spatial-resolution remote sensing (HSRRS) image classification tasks with their unique feature extraction and feature expression capabilities. However, the CNN-based classification method is very limited due to the acquisition of HSRRS images is difficult and the sample size is limited. In addition, the extraction of features by a single model is very limited, which limits the further improvement of classification performance. To solve the above problems, we propose ResNet50-InceptionV3 based on deep transfer learning and multi-feature fusion (TLMFFRI) model to apply for high-spatial-resolution remote sensing image classification. First, both ResNet50 and InceptionV3 are trained on the ImageNet dataset. Then, transfer the trained convolutional layers weights to the TLMFFRI model to fuse the features and realize the HSRRS image classification. Finally, we evaluate the method on the HSRRS dataset. Compared with ResNet50 based on transfer learning (TL-ResNet50) and InceptionV3 based on transfer learning (TL-InceptionV3), the proposed method achieved better classification performance.

With the ubiquitous network applications and the continuous development of network attack technology, all social circles have paid close attention to the cyberspace security. Intrusion detection systems (IDS) plays a very important role in ensuring computer and communication systems security. Recently, deep learning has achieved a great success in the field of intrusion detection. However, the high computational complexity poses a major hurdle for the practical deployment of DL-based models. In this paper, we propose a novel approach based on a lightweight deep neural network (LNN) for IDS. We design a lightweight unit that can fully extract data features while reducing the computational burden by expanding and compressing feature maps. In addition, we use inverse residual structure and channel shuffle operation to achieve more effective training. Experiment results show that our proposed model for intrusion detection not only reduces the computational cost by 61.99% and the model size by 58.84%, but also achieves satisfactory accuracy and detection rate.