Recent advancements in pretrained foundation models have shown considerable promise across various machine learning tasks. However, their application in the broader IoT industry remains limited, particularly in IoT physical-layer authentication. In this domain, the presence of domain shift between training and deployment environments, combined with privacy and data security concerns, renders traditional domain adaptation methods that rely on source domain data impractical. Motivated by these challenges, source-free unsupervised domain adaptation (SFUDA) presents a more feasible solution. In this paper, we propose a novel SFUDA framework that leverages a pretrained generative foundation model to augment target domain data without requiring access to source domain information. Additionally, we integrate an uncertainty-aware pseudo-labeling strategy along with consistency regularization to further enhance the adaptation process. Experimental results validate that our approach significantly outperforms conventional techniques, providing an effective and robust solution for IoT physical-layer authentication under realistic constraints.

Abstract— Radio Frequency Fingerprinting Identification (RFFI) serves as a pivotal technology in the Industrial Internet of Things (IIoT), witnessing significant strides over the last decade primarily due to advances in deep learning. However, most existing studies assume that training and test data are independent and identically distributed (i.i.d.), an assumption that often breaks down in real-world IIoT applications, leading to substantial performance degradation in cross-domain scenarios. To address this, we propose FATransformer, a novel unsupervised domain adaptation technique. The method is built on a robust theoretical foundation, ensuring stability and reliability across different domains. Specifically, FATransformer employs a Transformer-based architecture to efficiently process and align intermediate feature maps from both domains. Incorporating an attention-based module, it dynamically adjusts the weights of alignment at various layers, thereby improving the model’s flexibility to handle diverse real-world data. Extensive evaluations across multiple datasets underscore FATransformer’s superiority over existing methods.

Radio-frequency (RF) signal processing has seen significant advancements with the advent of deep learning, providing more accurate and efficient solutions for tasks such as signal classification and generation. However, most existing methods are heavily dependent on large labeled datasets, which are often scarce and costly to obtain in real-world RF environments. Furthermore, these approaches tend to be task-specific, limiting their ability to generalize across various RF applications. To address these challenges, this paper proposes RF-MAE, a self-supervised adaptive frequency masked autoencoder. RF-MAE leverages self-supervised learning (SSL) to capture intrinsic patterns from large-scale unlabeled RF data. Central to RF-MAE is a novel Adaptive Frequency Masked (AFM) strategy, which dynamically masks frequency components based on their energy distribution. Supported by a robust theoretical foundation, AFM ensures the model focuses on the most informative signal components, thereby enhancing generalization across RF tasks. By pretraining on unlabeled data and fine-tuning on specific tasks, RF-MAE significantly reduces the reliance on labeled datasets while improving adaptability across diverse RF signal processing tasks. Experimental results demonstrate that RF-MAE consistently outperforms traditional models, underscoring its potential to generalize across tasks and deliver superior performance in a wide range of RF signal applications.

Source-free unsupervised domain adaptation (SFUDA) aims to improve performance in unlabeled target domain data without accessing source domain data. This is crucial in scenarios with data-sharing restrictions due to privacy or compliance constraints. Existing SFUDA approaches often rely on pseudo-labeling techniques based on entropy or confidence metrics. These often overlook fine-grained data features, resulting in noisy pseudo-labels that degrade model performance. To overcome this limitation, we develop a new method called fine-grained pseudo-labeling and feature alignment (FGPLFA) to enhance SFUDA's performance. FGPLFA starts with a gradient-based metric that integrates insights from both model knowledge and data features, creating a more reliable sample metric. To enhance fine granularity, the fine-grained pseudo-labeling (FGPL) module was introduced. This module clusters data based on the magnitude and direction of gradients, allowing for dataset partitioning into subsets at the sample level. The subsets are pseudo-labeled with category-specificity and domain specificity, establishing a multilevel granularity structure that reduces noisy pseudo-labels. Subsequently, the mean-covariance adjustment feature alignment (MCAFA) method was introduced. Features from the subsets are aligned in a specified sequence, enhancing model adaptability in the target domain. Extensive experiments conducted across multiple datasets validate the superiority of FGPLFA.

Radio Frequency Fingerprinting Identification (RFFI) has become a critical technology in the physical-layer security (PLS) field, with deep learning emerging as the dominant approach over the past decade. However, most deep learning-based models rely on the assumption that training and testing data follow an independent and identical distribution (i.i.d.), which often does not hold in real-world scenarios. This mismatch significantly degrades model performance in cross-domain settings, making cross-domain RFFI a challenging task. Traditional unsupervised domain adaptation (UDA) methods attempt to address this issue by jointly optimizing task loss and domain loss which is able to reduce the distribution gap between training and testing data. However, we observe that during training, the gradients of these two losses often conflict, hindering effective optimization and limiting cross-domain performance improvements. To address these challenges, we propose a novel framework, Gradient Collaborative Optimization with Dynamic Weighted Feature Alignment (GCODWFA). Specifically, GCODWFA introduces a novel Gradient Collaborative Optimization (GCO) loss, which explicitly adjusts the gradient interaction between task and domain losses by optimizing their angular relationship. Additionally, it incorporates a Dynamic Weighted Feature Alignment (DWFA) strategy, which dynamically adjusts the layer-specific weights for feature alignment based on the angular similarity of task and domain gradients. Extensive experiments conducted on multiple datasets demonstrate the superiority of GCODWFA over existing methods.

Driven by the pursuit of enhanced performance, deep learning has recently seen rapid developments in the scaling of network architectures and parameters. However, this advancement has led to extremely high computational costs, undesirable in real-time and resource-limited scenarios. To address these challenges, we propose SwiftNet, a cost-efficient deep learning framework. Our novelty lies in SwiftNet's innovative multidimensional early-exit strategy that integrates seamlessly with existing neural network architectures. The framework includes additional branch classifiers concatenated to the backbone network, allowing high-confidence samples to exit early, thereby, reducing computational load. Unlike traditional methods, SwiftNet dynamically assesses confidence levels, ensuring only low-confidence samples proceed to subsequent classifiers or the final layer, optimizing resource usage without compromising accuracy. We have validated SwiftNet on multiple neural network models and datasets, demonstrating its ability to significantly reduce the computational cost of models while maintaining neural network performance.

While the Internet of Things (IoT) technology is booming and offers huge opportunities for information exchange, it also faces unprecedented security challenges. As an important complement to the physical-layer security technologies for IoT, radio frequency fingerprinting (RFF) is of great interest due to its difficulty in counterfeiting. Recently, many machine learning (ML)-based RFF algorithms have emerged. In particular, deep learning (DL) has shown great benefits in automatically extracting complex and subtle features from raw data with high-classification accuracy. However, DL algorithms face the computational cost problem as the difficulty of the RFF task and the size of the deep neural network have increased dramatically. To address the above challenge, this article proposes a novel cost-effective early-exit neural network consisting of a complex-valued neural network (CVNN) backbone with multiple random forest branches, called hybrid CVNN-RF. Unlike conventional studies that use a single fixed DL model to process all radio frequency (RF) samples, our hybrid CVNN-RF considers differences in the recognition difficulty of RF samples and introduces an early-exit mechanism to dynamically process the samples. When processing “easy” samples that can be well classified with high confidence, the hybrid CVNN-RF can end early at the random forest branch to reduce computational cost. Conversely, subsequent network layers will be activated to ensure accuracy. To further improve the early-exit rate, an automated multidimensional early-exit strategy is proposed to achieve scheduling control from multiple dimensions within the network depth and classification category. Finally, our experiments on the public ADS-B data set show that the proposed algorithm can reduce the computational cost by 83% while improving the accuracy by 1.6% under a classification task with 100 categories.

A prevailing assumption in existing deep learning research posits that data across source and target domains adhere to the independent and identically distributed (i.i.d.) assumption. However, this assumption often proves inadequate in real-world scenarios, leading to significant performance degradation when models encounter data with divergent distributions. To address this challenge, a novel unsupervised domain adaptation (UDA) algorithm, decoupling feature alignment (DFA), is introduced. The approach begins with the establishment of a robust theoretical framework, serving as the foundation for the mean-covariance adjustment feature alignment (MCAFA) algorithm. Simultaneously, a data decoupling (DD) module is introduced, effectively segregating target domain data into two subsets: one that mirrors the source domain and another that diverges markedly. Furthermore, a multidimensional alignment module is employed, leveraging the MCAFA algorithm and the DD module to align target data with source data across various layers and categories. Comprehensive evaluations on multiple data sets underscore the superiority of DFA.

Decentralized federated learning (DFL) has emerged as a conducive paradigm, facilitating a distributed privacy-preserving data collaboration mode in mobile edge computing (MEC) systems to bolster the expansion of artificial intelligence applications. Nevertheless, the dynamic wireless environment and the heterogeneity among collaborating nodes, characterized by skewed datasets and uneven capabilities, present substantial challenges for efficient DFL model training in MEC systems. Consequently, the design of an efficient collaboration strategy becomes essential to facilitate practical distributed knowledge sharing and cost reduction for MEC. In this paper, we propose an adaptive decentralized federated learning framework that enables heterogeneous nodes to learn tailored collaboration strategies, thereby maximizing the efficiency of the DFL training process in collaborative MEC systems. Specifically, we present an effective option critic-based collaboration strategy learning (OCSL) mechanism by decomposing the collaboration strategy model into two sub-strategies: local training strategy and resource scheduling strategy. In addressing inherent issues such as large-scale action space and overestimation in collaboration strategy learning, we introduce the option framework and a dual critic network-based approximation method within the OCSL design. We theoretically prove that the learned collaboration strategy achieves the Nash equilibrium. Extensive numerical results demonstrate the effectiveness of the proposed method in comparison with existing baselines.

With the development of wireless communication technology and the increasingly complex electromagnetic environment, radio frequency fingerprinting (RFF) plays a vital role in improving the security of communication and information systems. Recently, many RFF algorithms based on machine learning have emerged. However, most of them focus on improving the accuracy of RFF identification but ignore the computational cost. This paper proposes a novel classifier composed of a convolutional neural network (CNN) backbone with two random forest branches, called hybrid CNN-RF. Under the scheduling of the proposed multi-dimensional early exit strategy, hybrid CNN-RF can end early when processing “easy” samples to reduce the computational cost, and activate the inference of the subsequent network layer when processing “hard” samples to ensure accuracy. Finally, our experiments show that the proposed algorithm can reduce the computational cost by a factor of 3.41 while improving the accuracy by 1.5%.