Advanced communication systems and military reconnaissance are increasingly prevalent in high-tech environments, greatly supported by the flourishing in signal processing technologies. The recent exponential proliferation of sensors led to an unprecedented expansion in the scale and diversity of signals across various modalities. Such influx poses significant challenges in effectively integrating multi-modal signal data to deliver comprehensive and interpretive solutions across a diverse range of applications. In this paper, we provide an overview of the core issues, challenges, and future research directions in different stages of developing large-scale multi-modal signal processing models. Additionally, we introduce a prior investigation into signal representation learning, where we propose a contrastive learning-based framework to extract fine-grained signal features under few-shot conditions. Our proposed framework achieves a 24.1% performance improvement over baseline approaches, consistently demonstrating superiority over state-of-the-art methods. The code is accessible in this repository: https://github.com/YYH211/LSM .
With the rapid development of deep learning, its powerful feature extraction capabilities have shown great advantages in wireless signal modulation classification. However, limited by the inherent fragility of the model, deep learning-based modulation classification models are vulnerable to adversarial attacks. Consequently, defense frameworks integrating filtering mechanisms and adversarial learning have been developed. Yet, such approaches persistently confront the dual dilemmas of excessive feature degradation in signal processing or unwarranted dependence on prior knowledge assumptions, ultimately hindering real-world deployment. In this letter, we propose SigPurifier, a diffusion-based framework that eliminates adversarial perturbations in local data structures through forward diffusion while enhancing structural recovery through learned generative priors during signal denoising transitions. The evaluation results show that the proposed algorithm effectively eliminates adversarial attacks while maintaining signal communication quality. Compared to adversarial training methods, our model can increase the accuracy of the modulation classifier from 27.46% to 31.08%.
28 November 2025 in IEEE Wireless Communications Letters
As a fundamental security mechanism for wireless Internet of Things (IoT) devices, Radio Frequency Fingerprinting Identification (RFFI) faces significant challenges in open environments and cross-domain deployments. Among existing deep learning-based open-set recognition (OSR) methods, prototype-based approaches are widely adopted. However, these methods often suffer from prototype drift and feature misalignment between source and target domains due to channel effects, device load variations, and other factors. To address these challenges, this letter proposes a Prototype Calibration Domain Adaptation Framework guided by Pseudo-labeling and Domain-Adversarial Neural Networks (DANN), termed PCPD. The framework first leverages a prototype network to learn discriminative representations for each class in the source domain. Subsequently, it employs pseudo-label loss and a domain discriminator to dynamically calibrate domain-shifted prototype points, enhancing feature alignment across domains. Based on the calibrated prototypes, the method significantly improves OSR performance in cross-domain scenarios. Experimental results demonstrate that PCPD outperforms state-of-the-art algorithms, achieving an approximately 10.8% improvement in OSR accuracy under cross-domain conditions.
Existing zero-shot learning (ZSL) methods typically focus on mapping from the feature space (e.g., visual space) to class-level attributes, often leading to a non-injective projection. Such a mapping may cause a significant loss of instance-level information. While an ideal projection to instance-level attributes would be desirable, it can also be prohibitively expensive and thus impractical in many scenarios. In this work, we propose a variational disentangle zero-shot learning (VDZSL) framework that addresses this problem by constructing variational instance-specific attributes from a class-specific semantic latent distribution. Specifically, our approach disentangles each instance into class-specific attributes and the corresponding variant features. Unlike transductive ZSL, which assumes that unseen classes’ attributions are known beforehand, our VDZSL method does not rely on this strong assumption, making it more applicable in real-world scenarios. Extensive experiments conducted on three popular ZSL benchmark datasets (i.e., AwA2, CUB, and FLO) validate the effectiveness of our approach. In the conventional ZSL setting, our method demonstrates an improvement of 12∼15% relative to the advanced approaches and achieves a classification accuracy of 70% on the AwA2 dataset. Furthermore, under the more challenging generalized ZSL setting, our approach can gain an improvement of 5∼15% compared with the advanced methods.
Modulation recognition is critical in intelligent wireless communication, yet deep learning-based automatic modulation classification (AMC) models are vulnerable to adversarial attacks, posing severe risks. While adversarial detection and training offer partial mitigation, they suffer from evasion risks, signal distortion, or high latency—making them unfit for real-time systems like Unmanned Aerial Vehicle (UAV) swarms. Although generative models can purify adversarial inputs, their slow inference limits practicality. Conversely, reconstruction-based methods enable low-latency recovery but often compromise waveform fidelity. We propose a reconstruction-driven adversarial purification approach that directly restores clean signals at the input level, preserving both semantic features and physical consistency without classifier modification, ensuring high accuracy and real-time robustness. Experimental results on the RML2016.10b dataset show that our reconstruction-based method SigReconstruction achieves an average classification accuracy of 77.04% under adversarial attacks(clean accuracy of 86.68%). Reconstruction quality is corroborated by low mean squared erro (MSE) (0.0286/0.0031/0.0272) and low Fréchet Inception Distance (FID) (62.33/128.82/157.48), indicating faithful waveform recovery and feature alignment. These results demonstrate that targeted reconstruction with physical constraints offers practical, low-latency robustness for adversarially challenged wireless communications.
7 November 2026 in IEEE Transactions on Mobile Computing