基於深度學習之聯網自駕車切換車道異常偵測

Abstract

自動駕駛車輛可以利用各種感測器或或透過通訊系統來獲取周圍環境數據，以便在自主操控中做出決策。然而，感知或接收到的數據可能因惡意攻擊而導致錯誤。這對於具有高度安全性要求的自動駕駛車輛構成了嚴重威脅。在本文中，我們提出了基於深度學習的模型，用於偵測在具有切換車道意圖時是否受到攻擊，包括長短期記憶（LSTM）模型和深度神經網絡（DNN）模型。我們提出了兩種隱蔽攻擊模型，它們可以欺騙基於規則的偵測方法。我們在城市交通模擬軟體（SUMO）中直接部署這些攻擊，以生成異常數據。我們的異常檢測流程具有通用性，可應用於不同的變道環境，我們設計了三種環境進行實驗，包括高速公路、環狀交叉路口和對向超車。結果顯示，我們所提出的基於深度學習的方法對異常具有良好的偵測性能。

Autonomous vehicles can use various sensors or wireless networks to acquire their environmental data for making decisions in autonomous maneuvers. However, the sensed or received data can be malicious due to the attacks. This poses a serious threat to autonomous vehicles which are safety-critical systems. In this thesis, we propose deep-learning-based models, which are Long Short-Term Memory (LSTM) model and Deep Neural Network (DNN) model, to detect whether a vehicle is attacked while it has the lane-changing intention. We propose two stealthy attacks as attack models, which can deceive the detection by a rule-based detection approach. Then we directly deploy the attacks into Simulation of Urban Mobility (SUMO) during the simulation to generate the anomalous data. We also establish the standards and specifications for modifying simulation inputs in SUMO. Our anomaly detection workflow has the generality that can be used in different lane-changing environments, we design three environments to conduct experiments, including highway, roundabout, and opposite overtaking. As a result, the proposed deep-learning-based approach achieves a decent detection performance against the anomaly.