MODELESS: MODulation rEcognition with LimitEd SuperviSion
MODELESS: MODulation rEcognition with LimitEd SuperviSion
复制标题
DOI：
10.1109/secon52354.2021.9491617
复制
发表时间：
2021-07
期刊：
影响因子：
--
通讯作者：
Wei Xiong;Petko Bogdanov;M. Zheleva
中科院分区：
文献类型：
--
作者：
Wei Xiong;Petko Bogdanov;M. Zheleva
调制识别（modrec）是一项重要的发射机指纹识别任务，可以实现未来的频谱共享应用，如访问管理和强制执行。传统的监督模型需要标记所有目标调制的训练数据，这不能很容易地满足新的、定制的和数据驱动的波形的出现。因此，modrec适用性的一个关键问题是：我们能否通过适应和重用在不同但相关的调制上训练的模型来执行先前未观察到的调制的自动识别？为此，我们开发了MODELESS（具有有限监督的调制识别），它利用观察到的调制的知识对未观察到的样本进行分类。我们的解决方案基于零采样迁移学习，它利用观察和未观察类之间的侧信息来迁移学习过的分类器。特别是，我们量化了未观察到的和观察到的调制的理论星座图之间的相似性，并将它们应用于零射击迁移学习框架中。我们的框架是通用的，因为它可以产生任意调制的预测，只要它们的理论星座可以指定。我们在合成和现实世界的痕迹上评估MODELESS，并与文献中的零射击对应物进行比较。我们在大多数测试用例中展示了接近理想的分类准确性，并为未来对性能低于标准的分类任务的研究提出了建议。
Modulation recognition (modrec) is an essential transmitter fingerprinting task that enables future spectrum-sharing applications such as access management and enforcement. Traditional supervised modrec requires labeled training data for all target modulations, which cannot be readily met with the advent of new, customized and data-driven waveforms. Thus, a keystone question for the applicability of modrec is: Can we perform automatic recognition of previously unobserved modulations by adapting and reusing models that were trained on different but related modulations?To this end, we develop MODELESS (MODulation rEcognition with LimitEd SuperviSion) that exploits knowledge from observed modulations to classify samples from unobserved ones. Our solution is grounded in zero-shot transfer learning, which employs side information among observed and unobserved classes to transfer learned classifiers. In particular we quantify the similarity among the theoretical constellation diagrams of unobserved and observed modulations and employ them in a zero-shot transfer learning framework. Our framework is general, as it can produce predictions for arbitrary modulations as long as their theoretical constellations can be specified. We evaluate MODELESS on synthetic and real-world traces and in comparison with zero-shot counterparts from the literature. We demonstrate near-ideal classification accuracy in the majority of the testing cases and draw recommendations for future research into classification tasks with sub-par performance.