CRII: CNS: RUI: Exploiting Robust Deep Learning Framework for Wireless Localization Systems in Adversarial IoT Environments

随着无线网络和移动设备的激增，无线物联网(IoT)应用(如基于位置的服务)得到了相当大的关注。室内定位在无线传播环境中面临着许多挑战，包括多径效应、阴影、衰落和延迟失真。针对非视距(NLOS)室内环境，提出了一种基于指纹的深度神经网络无线定位方法。然而，使用DNN的纯数据驱动的方法在对抗性物联网环境(例如，无线干扰)中可能表现不佳。具体地说，DNN模型被证明容易受到通过引入细微扰动而产生的对抗性例子的攻击。因此，提出的研究的主要目的是为对抗性物联网环境下的无线定位开发健壮的解决方案，填补无线定位精度和健壮性之间的空白。特别是，我们考虑了对抗性机器学习在物联网环境中的无线定位。该项目的成功完成将显著提高无线本地化的最新水平，并实现强大的物联网应用。该项目的教育计划包括开发一门新的研究生级别的无线物联网系统深度学习课程，并加强各种核心本科和研究生级别的课程。此外，该项目努力扩大代表性不足群体对研究的参与，并将在整个项目年度继续大力加强这种努力。项目研究议程由两个紧密结合的研究主题组成。在推力I中，该项目将使用对抗性深度学习进行室内本地化，其方式是利用离线阶段的对抗性训练来提高深层网络的健壮性，从而减轻对抗性示例攻击对无线数据的威胁。这个项目将考虑两个无线定位任务：黑盒攻击中无线定位的对抗性范例和对抗性范例检测的无监督学习。在推力II中，该项目将结合深度学习和高斯过程进行不确定位置估计，以提高无线定位算法的稳健性。具体地说，该项目将利用深度高斯过程的不确定性位置估计来对抗白盒和黑盒攻击。此外，本项目还将对无线定位的基本限制和健壮性进行建模和分析。对于两次推进中提出的所有任务，该项目将开发数学模型和求解算法。建议的算法将在无线物联网设备/平台(例如Wi-Fi、RFID和LoRa)上实施，并在具有代表性的室内环境中进行广泛的实验验证。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

With the proliferation of wireless networks and mobile devices, wireless Internet of Things (IoT) applications (e.g., location-based services) have gained considerable attention. Indoor localization faces a number of challenges in the radio propagation environment, including the multipath effect, shadowing, fading, and delay distortion. To tackle the non-line-of-sight (NLOS) indoor environment, fingerprinting based wireless localization methods using deep neural networks (DNN) have been proposed. However, a data-driven only approach using DNN may perform poorly in adversarial IoT environments (e.g., wireless jamming). Specifically, DNN models are shown to be vulnerable to adversarial examples generated by introducing a subtle perturbation. Thus, the primary aim of the proposed research is to develop robust solutions for wireless localization in adversarial IoT environments, which fills in the gap between wireless localization accuracy and robustness. Particularly, we consider adversarial machine learning for wireless localization in IoT environments. The successful completion of this project will significantly improve the state-of-the-art of wireless localization and enable robust IoT applications. The project's educational plan includes developing a new graduate-level course on deep learning for wireless IoT systems and enhancing various core undergraduate and graduate-level courses. Also, the project strives to broaden participation from under-represented groups in research and will continue to greatly strengthen such efforts throughout the project years.The project research agenda is composed of two closely integrated research thrusts. In Thrust I, this project will use adversarial deep learning for indoor localization in a way that leverages adversarial training in the offline stage to improve the robustness of the deep network, thus alleviating the threat of the adversarial example attacks on wireless data. This project will consider two wireless localization tasks: adversarial examples for wireless localization in black-box attacks and unsupervised learning for adversarial examples detection. In Thrust II, this project will combine deep learning and Gaussian processes for uncertain location estimation, to improve robustness for wireless localization algorithms. Specifically, this project will exploit uncertainty location estimation with deep Gaussian process against both white-box and black-box attacks. Also, this project will model and analyze the fundamental limits and robustness of wireless localization. For all the proposed tasks in the two thrusts, this project will develop mathematical models and solution algorithms. The proposed algorithms will be implemented with wireless IoT devices/platforms (e.g., Wi-Fi, RFID, and LoRa), and validated with extensive experiments in representative indoor environments.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.