Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/57383
Title: | 基於深度學習的協同式自適應巡航控制系統侵入偵測 Deep-Learning-Based Intrusion Detection for Cooperative Adaptive Cruise Control |
Authors: | Sing-Yao Wu 吳星耀 |
Advisor: | 林忠緯(Chung-Wei Lin) |
Keyword: | 協同式自適應巡航控制系統,車聯網,異常偵測,侵入偵測,機器學習,深度學習, Cooperative Adaptive Cruise Control (CACC),VANET,Anomaly Detection,Intrusion Detection,Machine Learning,Deep Learning, |
Publication Year : | 2020 |
Degree: | 碩士 |
Abstract: | 協同式自適應巡航控制系統(CACC)為先進駕駛輔助系統(ADAS)中的一環,旨在提升車子跟車時的安全、效率及車串的穩定性。協同式自適應巡航控制系統利用車上安裝的感測器及車聯網的訊息交換,獲得諸如前車的位置、速度、加速度等資訊,藉此更加安全的控制與前車的車距。然而,這些由感測器與車聯網所獲得的資訊具有被攻擊者侵入的風險,一旦遭受侵入,攻擊者便可透過修改這些資訊使系統無法準確保持車距,造成損害。 為了應對這個問題,本研究我們考慮三種針對CACC的攻擊:脈衝攻擊、線性攻擊與隱匿攻擊。其中,我們特別設計了隱匿攻擊來欺騙一個規則式偵測法。接著,我們提出三個不需要攻擊模型作訓練且基於深度學習的侵入偵測方法:預測式偵測法、編碼解碼式偵測法與對抗生成網路式偵測法來檢測協同式自適應巡航控制系統的輸入是否被惡意修改。實驗結果展現了各偵測方法的優勢與劣勢,且顯示在偵測所考慮的攻擊中,對抗生成網路式偵測法的偵測能力最穩定。 Cooperative Adaptive Cruise Control (CACC) is an advanced system to improve safety, efficiency, and string stability for a vehicle following its leading vehicle. The vehicle equipped with CACC utilizes vehicular communication and receives information such as the position, velocity, acceleration, or even intention from the leading vehicle. However, it has been shown that sensors and wireless channels are vulnerable to security attacks, and attackers can modify data sensed from sensors or sent from other vehicles. To address this problem, in this thesis, we consider three types of attacks, pulse attacks, linear attacks, and stealthy attacks, on CACC inputs. Especially, we design the stealthy attacks to deceive a rule-based approach. We then develop three deep-learning models, a predictor-based model, an encoder-decoder-based model, and a Generative Adversarial Network-based (GAN-based) model, to detect the attacks, where the three models do not need attacker models for training. We further apply dynamic thresholding for error measure to improve the detection performance. The experimental results demonstrate the respective strengths of different models and suggest that the GAN-based model has the most stable performance of detecting the attacks. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/57383 |
DOI: | 10.6342/NTU202001697 |
Fulltext Rights: | 有償授權 |
Appears in Collections: | 資訊工程學系 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
U0001-2107202016054900.pdf Restricted Access | 2.44 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.