Orthogonal Frequency Division Multiplexing (OFDM)-based waveforms are widely employed in current and emerging Internet of Things (IoT) communication systems. However, in practical deployments, IoT communications frequently encounter complex environmental factors such as sparse pilot allocation, non-linear hardware distortion, and interference, which significantly degrade the detection performance of traditional OFDM receivers. To address these challenges, this paper proposes a structured hybrid deep learning (DL) OFDM receiver architecture, termed DCSE-DAR, which integrates model-driven signal processing with data-driven feature learning. The proposed receiver comprises a Dilated Convolution with Squeeze-and-Excitation (DCSE) module for enhanced channel estimation and a Dual-Activation Residual (DAR) network that unfolds the Minimum Mean Square Error (MMSE) principle into a deep iterative structure. Through structured fusion between these modules and a collaborative two-stage training strategy, the proposed architecture enables effective feature extraction and model-guided signal detection while preserving the interpretability of the classical OFDM receiver pipeline. Simulation results show that the DCSE-DAR receiver is superior to the conventional MMSE receiver as well as representative DL-based receivers under various channel models, sparse pilot configuration, nonlinear distortion and interference conditions, showing better robustness and detection accuracy, which verifies its effectiveness and application potential in the design of intelligent receiver in time-varying IoT communication environment.

Specific emitter identification (SEI) is a security authentication technology by utilizing radio frequency fingerprinting (RFF) features. However, current mainstream RFF feature extraction methods based on neural networks (NNs) generally suffer from poor interpretability and difficulties in architecture optimization. To address these issues, we propose a novel convolution-enhanced swin vision transformer (CoSwinVIT) that combines filter characteristics of NN architectures to achieve a uniform spectrum response. Specifically, we treat the NN used in SEI tasks as filters and investigate their filter characteristics by applying Fourier transforms to hidden vectors. Through spectral analysis, we found that different NN architectures exhibit significantly different responses to the signal spectrum. This affects the model’s sensitivity to specific frequency bands of the signal, thereby influencing its accuracy (Acc). Subsequently, by integrating the filtering properties of swin vision transformer (SwinVIT) and convolutional neural network (CNN), we achieve a uniform spectral response design. Finally, to evaluate the performance of the CoSwinVIT architecture, we design both a supervised learning algorithm and a contrastive learning-based self-supervised algorithm. Experimental results on a real-world automatic dependent surveillance-broadcast (ADS-B) and wireless fidelity (WiFi) dataset indicate that CoSwinVIT provides a more uniform spectrum response. Under supervised learning, the proposed CoSwinVIT obtains the accuracies of 98.6% and 99.8% on the ADS-B and WiFi datasets, respectively. Under self-supervised learning with 5% and 10% labeled data, the accuracies of 73.6% and 86.3% are achieved on the ADS-B dataset, and the accuracies of 81.6% and 91.5% are achieved on the WiFi dataset. These results surpass the state-of-the-art (SOTA) methods used in the SEI tasks.

Deep learning has demonstrated promising results over traditional hand-crafted methods for automatic modulation classification (AMC), which plays a critical role as an intermediate step between signal detection and modulation. However, acquiring large-scale labeled data remains challenging, as both data quality and annotation costs are critical factors in achieving accurate and efficient training. In this paper, we propose a novel self-supervised contrastive learning (SSCL) with gradient-adversarial-based data augmentation (GADA) approach for AMC. Additionally, a meticulous encoder based on Transformer-LSTM architectures is employed to pre-train a feature extractor using unlabeled base classes. Subsequently, knowledge transfer is employed to fine-tune the feature extractor, and ensemble learning is introduced to efficiently leverage multiple classifiers for joint decision-making. Experiments demonstrate that we achieve up to 91.87% accuracy on the challenging large-scale RadioML2018.10a dataset, demonstrating performance competitive with state-of-the-art supervised implementations.

The broadcast nature of device authentication systems in the Internet of Things (IoT) and maritime transportation makes real-time signals susceptible to forgery attacks, leading to critical risks, such as information leakage and navigation failures. specific emitter identification (SEI) plays a crucial role in protecting communication systems from spoofing and tampering. However, traditional SEI networks struggle with limited generalizability, high-computational costs, and susceptibility to extraction and inversion attacks. To overcome these challenges, this article introduces sliding window linear attention and the convolution mechanism (SWLC)-Conformer, a novel hybrid SEI architecture that seamlessly integrates SWLC. The proposed architecture dynamically captures both local signal characteristics and global temporal dependencies through a hybrid design incorporating window-based attention blocks and depth-wise convolution layers. Furthermore, to strengthen system security against model inversion threats, we incorporate a perturbation mechanism that employs reverse sigmoid activation in the final feature representation. Experimental results demonstrate that SWLC-Conformer achieves high recognition accuracy, reaching 91.10% on the AIS100 dataset and 97.89% on the WiSig dataset, while maintaining FLOPs at 22.17M and 11.11M, respectively. In terms of security, SWLC-Conformer reduces the accuracy of stolen models by over 40% when the stolen dataset ratio is 0.9. This result highlights the effectiveness of the introduced defense mechanism in degrading the usability of stolen models through controlled feature obfuscation.

Specific Emitter Identification (SEI) is a non-encrypted authentication technology that adds a layer of security to wireless communications in intelligent transportation systems. However, most existing SEI methods are constrained to identify fixed classes and cannot learn incrementally. In practical applications, the constant emergence of new classes or tasks necessitates the capacity to continuously learn new classes from streaming data. In this paper, combining momentum-based prototype correction and hierarchical regularization, a simple and effective non-example class-incremental method for SEI is proposed, named MoPC-HR. The momentum-based prototype correction dynamically adjusts old class prototypes (i.e., old class centers), strengthening the separation between old and new classes and enabling smoother integration of new classes. Hierarchical regularization is applied at multiple levels to control feature differences between old and new classes, preventing the model from favoring new classes and reducing catastrophic forgetting. Experiments on real-world AIS-100 and ADS-B datasets show that MoPC-HR outperforms state-of-the-art methods. Specifically, in the incremental phase 20, MoPC-HR achieves an average accuracy of 97.04%, average forgetting of 2.95%, and average intransigence of 2.85% on the AIS-100 dataset; and an average accuracy of 95.99%, average forgetting of 3.66%, and average intransigence of 3.60% on the ADS-B dataset. The code is available at https://github.com/xmuLdz/MoPC-HR.git