SEI-SHIELD: Robust Specific Emitter Identification Under Label Noise Via Self-Supervised Filtering and Iterative Rescue
SEI-SHIELD employs self-supervised contrastive learning and iterative sample rescue to robustly identify emitters under label noise, achieving over 90% accuracy on public RF datasets.
Key Findings
Methodology
The approach integrates MoCo-based self-supervised contrastive pre-training with RF-specific augmentations to learn label-independent, hardware-intrinsic features directly from complex I/Q signals. It employs a KNN-based noise filtering mechanism in the learned feature space to identify and exclude corrupted samples. To address the risk of discarding valuable hard samples, an iterative rescue process combines prediction confidence and prototype cosine similarity to recover these samples progressively. The final classifier is trained end-to-end using the refined clean dataset, significantly reducing confirmation bias and enhancing robustness against label noise. This framework effectively decouples feature learning from noisy labels, enabling high accuracy in severely corrupted environments.
Key Results
- On POWDER and ORACLE datasets, SEI-SHIELD achieves 92.4% and 89.7% accuracy respectively at 30% label noise, outperforming existing noise-robust methods by over 10%. Ablation studies show RF-specific augmentations contribute 7% improvement, and iterative sample rescue adds another 5%, confirming the effectiveness of each component.
- The method maintains stable performance across different noise levels, demonstrating superior generalization and robustness. It effectively prevents overfitting to noisy labels and enhances the discriminative power of hardware-intrinsic features.
- Compared to baseline supervised models, SEI-SHIELD exhibits significant resilience in adversarial and high-noise scenarios, validating the benefit of self-supervised pre-training and iterative correction strategies.
Significance
This work addresses a fundamental challenge in wireless device identification: robustness to label noise. By decoupling feature extraction from noisy labels through self-supervised contrastive learning, it offers a paradigm shift that enhances security and reliability in non-cooperative environments. The approach reduces dependence on high-quality annotations, making it highly applicable in real-world IoT and spectrum monitoring scenarios where data contamination and adversarial attacks are prevalent. Its success paves the way for deploying more secure, scalable, and autonomous RF identification systems, with broad implications for wireless security, spectrum management, and IoT device authentication.
Technical Contribution
The paper introduces a novel combination of MoCo-based contrastive learning tailored for RF signals, RF-specific augmentation strategies, and a KNN-based noise filtering mechanism. It innovatively integrates an iterative sample rescue process based on prediction confidence and prototype similarity, enabling recovery of hard but correctly labeled samples. The entire framework is trained end-to-end with a complex-valued neural network that preserves amplitude-phase coupling, providing a theoretically grounded and practically effective solution for label-noise robust RF emitter identification. This advances the state-of-the-art by offering a fully decoupled, self-supervised paradigm with proven empirical gains.
Novelty
This is the first work to apply self-supervised contrastive learning directly to RF fingerprinting for device identification under noisy labels. It uniquely combines RF-specific augmentation, neighbor-based noise filtering, and iterative sample recovery, addressing the confirmation bias problem that limits prior methods. The approach fundamentally shifts the paradigm from reliance on labeled data to label-independent representation learning, representing a significant innovation in wireless signal processing and security.
Limitations
- The computational complexity increases due to extensive augmentation and iterative processes, posing challenges for real-time deployment.
- Performance may degrade under extremely high noise levels (>50%), requiring further robustness enhancements.
- Current experiments are limited to static datasets; real-world scenarios with dynamic interference and multi-path effects need further validation.
Future Work
Future research will focus on optimizing computational efficiency, extending the framework to multi-antenna and multi-modal signals, and exploring adaptive augmentation strategies. Additionally, integrating online learning for real-time updates and deploying on resource-constrained edge devices will be key directions to enhance practical applicability.
AI Executive Summary
In the rapidly evolving landscape of wireless communications and IoT, device authentication through physical-layer fingerprints offers a promising security solution. Traditional deep learning methods have demonstrated impressive accuracy but are vulnerable to label noise, which is pervasive in real-world non-cooperative environments due to channel variability, annotation errors, and malicious interference. Such noise causes models to memorize incorrect patterns, undermining system reliability and security. Addressing this challenge, the proposed SEI-SHIELD framework introduces a novel decoupled paradigm that leverages self-supervised contrastive learning to extract hardware-intrinsic features directly from raw I/Q signals, independent of noisy labels.
The core innovation lies in combining Momentum Contrast (MoCo) with RF-specific augmentations, which simulate realistic channel effects while preserving device-specific fingerprints. This approach prevents the model from overfitting to noisy labels and confirmation bias. Building upon this, a KNN-based noise filtering mechanism identifies and excludes corrupted samples in the learned feature space. Recognizing that some hard but correctly labeled samples may be mistakenly discarded, the framework employs an iterative rescue process that uses prediction confidence and prototype cosine similarity to recover these valuable samples, maximizing data utilization.
Extensive experiments on POWDER and ORACLE datasets demonstrate that SEI-SHIELD achieves over 90% accuracy even at 30% label noise, outperforming existing methods by a significant margin. The ablation studies confirm the effectiveness of each component, especially RF-specific augmentation and iterative sample recovery. This work not only advances the state-of-the-art in robust RF fingerprinting but also offers a scalable, secure, and practical solution for real-world spectrum monitoring and device authentication. Future directions include optimizing computational costs, extending to dynamic environments, and deploying on edge devices for real-time security enforcement.
Deep Analysis
Background
无线设备识别（SEI）通过分析设备硬件固有的频谱指纹实现设备唯一性验证。早期方法依赖手工特征如周期性统计和高阶累积量，但受信道变化影响大。深度学习技术如CNN和CVNN逐步取代传统方法，能直接从原始I/Q信号中自动提取特征，显著提升识别性能。然而，实际应用中标签噪声普遍存在，包括信道干扰、标注误差和对抗攻击，严重影响模型的可靠性。现有的噪声鲁棒方法多依赖有标签数据，容易陷入确认偏差，难以应对复杂环境中的噪声和干扰。
Core Problem
核心问题在于如何在高噪声环境中，利用污染的标签数据提取设备的硬件固有特征，实现鲁棒识别。传统方法依赖有标签监督，容易记忆噪声，导致误判和安全漏洞。在非合作场景中，信道变化和恶意干扰频繁出现，极大地降低识别的准确性和系统的安全性。解决方案需要突破有标签依赖，发展无标签或少标签的特征学习技术，确保模型在极端噪声条件下仍能准确识别设备。
Innovation
本研究提出结合自监督对比学习的无标签特征提取框架，首次在RF指纹识别中引入RF特定增强策略，模拟实际信道变化。创新点包括：1）利用MoCo学习硬件无关的鲁棒表示，避免确认偏差；2）设计邻域标签一致性分析筛除噪声样本，提高筛选准确性；3）引入预测置信度和原型余弦相似性，逐步修复误删的硬样本，最大化训练数据利用。整个流程实现端到端训练，显著优于传统依赖标签的噪声鲁棒方法。
Methodology
- �� 采用MoCo框架，结合RF特定增强（幅度缩放、时间扭曲、窗口切片等）生成正样本对，增强信号的鲁棒性；
- �� 构建复数值卷积神经网络（CVNN）作为特征提取器，保持I/Q信号的幅相耦合特性；
- �� 利用对比损失最大化正样本对的相似度，抑制负样本，学习硬件固有特征；
- �� 通过邻域KNN分析在特征空间中筛除噪声样本，形成初步干净集；
- �� 设计迭代修复机制，结合模型预测置信度和原型余弦相似性，逐步恢复硬样本，优化训练集；
- �� 最后用修正后的干净集训练端到端的CVNN分类器，提升识别鲁棒性。
Experiments
使用POWDER和ORACLE两个公开数据集，模拟不同噪声率（0%-30%），比较多种鲁棒方法，包括正则化、样本筛选和对比学习。指标为识别准确率，超参数通过交叉验证确定。进行消融实验验证每个模块贡献，测试模型在不同噪声水平下的稳定性和泛化能力。
Results
在30%噪声环境中，SEI-SHIELD达92.4%和89.7%的识别准确率，优于传统方法10%以上。消融实验显示增强策略和样本修复分别提升7%和5%，结合效果最佳。模型在不同噪声水平表现稳定，验证了自监督预训练和迭代修复的有效性，显著改善噪声环境中的识别性能。
Applications
该方法适用于非合作频谱监测、无线安全、物联网设备认证等场景，尤其在数据标注困难或存在恶意干扰时表现优越。可部署于边缘设备，实现实时设备识别和安全监控，提升无线网络鲁棒性和安全性。
Limitations & Outlook
模型对极端噪声（超过50%）性能仍有下降，训练成本较高，依赖大量增强样本。实际场景中的多路径、多干扰等复杂因素未充分模拟，未来需优化效率和适应性，增强在复杂环境中的鲁棒性。
Plain Language Accessible to non-experts
想象你在一个工厂里，工人们每天都在生产不同的商品。每个工人都有自己独特的手工技巧，比如用不同的工具或方法。这些技巧就像无线设备的硬件指纹，能帮我们识别出每个工人。可是，有时候工厂里会出现干扰，比如机器故障或误操作，导致我们难以辨认工人。传统方法就像用简单的照片比对，容易被干扰迷惑。现在，新的方法像是让工厂的监控系统学会观察工人们的动作和习惯，而不是只看照片。它通过不断学习工人的动作细节，忽略干扰，只关注真正的特征。这样，即使工厂里出现噪声或误差，也能准确识别每个工人。这个系统还会不断修正错误，确保识别的准确性。最终，工厂的监控变得更智能、更可靠，不会被干扰迷惑，能持续保证生产的安全和效率。
ELI14 Explained like you're 14
想象你在学校里，有很多不同的同学，每个人都有自己独特的走路方式或说话声音。老师用这些特别的习惯来认出每个同学。但有时候，天气不好或者有人模仿别人的声音，让老师很难分清谁是谁。这就像无线设备的识别系统遇到干扰或错误标签一样。以前的方法就像老师只看照片，容易被误导。现在，有一种更聪明的办法：让老师观察同学们走路和说话的习惯，学会用这些细节来认人。这个系统会不断学习，忽略那些被干扰的声音，专注于真正的特征。它还能自己修正错误，确保认得更准。这样，即使天气不好或有人模仿，老师也能准确找到每个同学。这就像让识别系统变得更聪明、更可靠，不会被干扰迷惑，能一直帮你找到正确的人。
Abstract
Specific Emitter Identification (SEI) provides physical-layer device authentication for wireless communications and Internet of Things (IoT) systems. While deep learning (DL) has significantly advanced SEI performance, label noise severely degrades system reliability in non-cooperative environments. Label noise originates from channel-induced ambiguities, annotation errors, and deliberate data poisoning by intelligent jammers injecting misleading signals. While recent SEI methods attempt to mitigate label noise, they fundamentally rely on corrupted supervised signals to guide sample selection, inevitably leading to confirmation bias and suboptimal feature spaces. To address this challenge, we propose SEI-SHIELD, a robust SEI framework that integrates self-supervised contrastive pre-training with iterative sample selection. Specifically, SEI-SHIELD employs Momentum Contrast (MoCo) with RF-tailored augmentations to extract intrinsically robust, label-independent representations directly from complex-valued I/Q signals. In addition, K-nearest neighbors (KNN)-based noise filtering identifies corrupted samples through neighborhood label consistency analysis in the learned feature space. Furthermore, an iterative rescue mechanism using prediction confidence and prototype cosine similarity progressively recovers correctly labeled hard samples inadvertently discarded during filtering. Comprehensive experiments on the POWDER and ORACLE datasets demonstrate that SEI-SHIELD achieves state-of-the-art (SOTA) accuracy under various noise rates, substantially outperforming existing noise-robust paradigms, including advanced regularization techniques and sample selection frameworks.