Decades of practices of ramp metering, by controlling downstream volume and smoothing the interweaving traffic, have proved that ramp metering can decrease total travel time, mitigate shockwaves, decrease rear-end collisions, reduce pollution, etc. Besides traditional methods like ALIENA algorithms, Deep Reinforcement Learning algorithms have been established recently to build finer control on ramp metering. However, those Deep Learning models may be venerable to adversarial attacks. Thus, it is important to investigate the robustness of those models under False Data Injection adversarial attack. Furthermore, algorithms capable of detecting anomaly data from clean data are the key to safeguard Deep Learning algorithm. In this study, an online algorithm that can distinguish adversarial data from clean data are tested. Results found that in most cases anomaly data can be distinguished from clean data, although their difference is too small to be manually distinguished by humans. In practice, whenever adversarial/hazardous data is detected, the system can fall back to a fixed control program, and experts should investigate the detectors status or security protocols afterwards before real damages happen.

翻译：匝道控制通过控制下游流量并优化交织车流，数十年的实践已证明其能够减少总行程时间、缓解冲击波、降低追尾事故及减少污染等。除ALIENA算法等传统方法外，近年来深度强化学习算法已被应用于实现更精细的匝道控制。然而，这些深度学习模型可能易受对抗性攻击。因此，研究这些模型在虚假数据注入对抗攻击下的鲁棒性至关重要。此外，能够从清洁数据中检测异常数据的算法是保障深度学习算法的关键。本研究测试了一种可在线区分对抗数据与清洁数据的算法。结果发现，在大多数情况下，异常数据能够与清洁数据区分开，尽管二者差异极小，人类难以通过手动方式辨别。实践中，一旦检测到对抗性/危险数据，系统可切换至固定控制程序，专家应在实际损害发生前检查探测器状态或安全协议。