Electronic and electromechanical devices continuously emit electromagnetic (EM) signals while in use. These EM emissions (EMEs) contain unique spectral characteristics that can be leveraged for a variety of purposes, including identification of the device's unique EM “fingerprint,” and characterization of software processes running on the device. In this study, we implemented a novel method for automatic identification and characterization of these EMEs inspired by a classification/decoding scheme used to extract neural correlates of brain state from magnetoencephalographic data. Utilizing a sparse bilinear formulation of logistic regression as our “neural” decoder, we extracted device- and software-specific spectrospatial patterns from five identical Arduino Uno prototyping boards as they cycled through five program states. In the device fingerprinting task, we were able to discriminate all five boards from each other with near-perfect accuracy using this method. For software characterization, within a single device, we detected all five programs with 90% accuracy, and across devices, we were able to identify 3/5 programs with 99% accuracy, and achieving 77% accuracy for the other two. Overall, this neural-decoding approach performed well in all scenarios tested, and the corresponding EME “maps” it provides quantify the differences between device and software-specific EMEs in an easily interpretable way.

中文翻译：

---

使用神经解码算法通过无意电子发射对电子设备和软件过程进行分类

电子和机电设备在使用过程中会持续发射电磁 (EM) 信号。这些 EM 发射 (EME) 包含独特的光谱特征，可用于多种用途，包括识别设备独特的 EM“指纹”，以及表征设备上运行的软件进程。在这项研究中，我们实施了一种自动识别和表征这些 EME 的新方法，其灵感来自用于从脑磁图数据中提取大脑状态的神经相关性的分类/解码方案。利用逻辑回归的稀疏双线性公式作为我们的“神经”解码器，我们从五个相同的 Arduino Uno 原型板中提取设备和软件特定的光谱空间模式，因为它们在五个程序状态中循环。在设备指纹识别任务中，我们能够使用这种方法以近乎完美的准确度将所有五块板相互区分。对于软件表征，在单个设备中，我们以 90% 的准确率检测到所有五个程序，在跨设备中，我们能够以 99% 的准确率识别 3/5 的程序，而其他两个程序的准确率达到 77%。总的来说，这种神经解码方法在所有测试场景中都表现良好，并且它提供的相应 EME“地图”以一种易于解释的方式量化了设备和软件特定 EME 之间的差异。我们能够以 99% 的准确率识别 3/5 的程序，并为其他两个程序实现 77% 的准确率。总的来说，这种神经解码方法在所有测试场景中都表现良好，并且它提供的相应 EME“地图”以一种易于解释的方式量化了设备和软件特定 EME 之间的差异。我们能够以 99% 的准确率识别 3/5 的程序，并为其他两个程序实现 77% 的准确率。总的来说，这种神经解码方法在所有测试场景中都表现良好，并且它提供的相应 EME“地图”以一种易于解释的方式量化了设备和软件特定 EME 之间的差异。