context-aware automotive intrusion detectionmpese/papers/escar17... · •adaptive...
TRANSCRIPT
Armin Wasicek1
Mert D.Pesé2, André Weimerskirch2, Yelizaveta Burakova2, Karan Singh2
1Technical University Vienna, Austria2University of Michigan
ESCAR USAJune 21, 2017
Context-aware Automotive Intrusion Detection
Motivation for Security
Dominant motives of organizations for security:
• Avoiding and mitigating loss
• Avoiding negligence
• Enhancing strategic business values
► Security is the enabler of business models
2Armin Wasicek
Who is willing to pay for Security?
Incentives for investment in security are lacking
• Making a strong business case for security is hard:
– Quantifying risk in a dynamic environment
– Rapidly changing, unpredictable threats
– Demonstrating consequences of attacks
►Who is willing to pay for security?
3Armin Wasicek
Vulnerability Black Market
Zero-day exploits are a tradable asset
• Full disclosure to improve software is utopia
• New professionalism in security market
• Cost of exploit: $ 1k-5k (2k), good >10 month
• “Realistically, we’re selling cyberweaponry”
► There is already a market for security!
4Armin Wasicek
Rising need for Security
Automotive security requirements
• Cost effective
• Reusable
• Adaptive
• Cyber-Physical nature
►Raise the bar to prevent (simple) attacks
Armin Wasicek 5
Security toolkit
ECU security
Secure On-board Comm.
Perimeter security
Intrusion detection
What is Intrusion Detection?
Gathers and analyzes information
• Identifies potential security breaches
– Intrusions
– Misuse/Fraud
► Reports to users
6
System
Perimeter
Users
Sensor
IDS
Armin Wasicek
Manipulation and Fault tolerance
7
• Triggering unsafe states will stop the system
• Manipulations are subtle
• Stay within safe states, but modified behavior
• Recognition and Recovery
Armin Wasicek
NTHSA: Misbehavior Detection [DOT HS 812 014]
Development of the processes, algorithms, reporting requirements, and data requirements for both local and global detection functions;
Types of IDS
• Knowledge-based IDS– Patterns/Signatures of malicious activities
– Low false positive rate, needs frequent updates
• Heuristic-based IDS– Look for abnormal behavior, e.g., higher entropy
– Detect new attack patterns
• Context-aware IDS– Compare to reference model, include semantics
– Check against specifications and regulations
8Armin Wasicek
IDSS
Automotive System Architecture
Armin Wasicek 9
ECU
Switch
BackboneCloud
Eth
GW
• Host-based IDS monitors ECU– CPU & memory usage, syscalls, # processes, …
• Network IDS monitors communication– Message frequency, patterns, entropy, …
Over-the-air UpdatesEnvironmental info.
Malicious devices
Board computer
External communication
On-board networks
Segregation
Identify anomalies and outliers
Chip tuning
Modify control algorithm parameters in ECU
• Parameters are stored in a table in flash memory
• Reprogram ECU with new values
– Debug interface, 3rd party device
►Messages emitted by ECU seem original!
10Armin Wasicek
Power boxing
Modify commands to ECU
• Replace the ECU in the communication system
• Insert device between the ECU and actuators
► Communication pattern does not change!
11
Improves low end torque. Plug-in installation in less than 30 minutes.
Armin Wasicek
Cyber-Physical Attacks
Automotive systems are Cyber-Physical
• Checking only cyber properties like CAN message frequency might miss important attack vectors
• IDS needs to target attack on the physical part
► Compare actual behavior to reference model enabling misbehavior detection
Armin Wasicek 12
ECU
Switch
Backbone
GW
Automotive System Architecture
Armin Wasicek 13
Cloud
• Integrate firewall, authentication, and detection
• Fuse information from diverse sources
• Use semantics of control msg to reason about manipulation
Threat defenseOver-the-air UpdatesDetect misbehavior
Chip Tuning
Wheel speed
RPM, torque
Road conditions
IDS
Feature Selection
Select parameters capturing engine behavior
Engine control measures many useful parameters
• Speed …velocity of vehicle
• RPM …angular velocity of engine
• Torque…turning force
►We want to detect modifications, like going around the curve slightly faster, having more torque on a slope, etc.
14
Intrusion Detection Layer
Compares current to reference behavior
• Trained model reconstructs some features poorly
• These are considered as outliers
►Manipulation, if num. of outliers exceeds threshold
15Armin Wasicek
Feature Extraction
Convert a time series to a feature vectorProcessing pipeline works on a time slice
• Use original signal and its differentiation
• Use statistical moments to expand features
• Normalize feature vector16Armin Wasicek
Solve a one-class classification problem
Autoencoder Neural Network:
• Hidden layer generalizes ratio between features
• Stores the typical behavior of an engine
• Trained using same vector for input X, output Y
• Anomaly score is error between input and output
Artificial Neural Networks
17Armin Wasicek
Evaluation
• Racing car simulation TORCS
• Car model ‘p406’ simulating a Peugeot 406
• Increase engine torque by 10 Nm and 30 Nm
• Let the robot do the driving!
18Armin Wasicek
Engine tuning
Modification
• Modify RPM/Torque ratio
Recognizing manipulation
• Bottleneck ANN
– 8 hidden perceptron
• ROC curve looks promising
Armin Wasicek 19
Note: This is unsupervised learning!
Experimental Setup
Armin Wasicek 20
Power boxing to manipulate engine control Read out
data via OBD II
Perform analytics, detect presence
Urban/Highway driving cycles
Real car data
Data points
0 5 10 15 20 25 30 35 40 45 50 55
Min
/max v
alu
es
0
5
10
15
20
25
30
35
40
45
original
modified
Armin Wasicek 21
Vehicle speed Calculated load valueEngine RPM Absolute throttle positionFuel rate O2 sensor lambda wide range Fuel/Air commanded equivalence Absolute throttle position B Accelerator pedal position D Catalyst temperature
Recognition result
Armin Wasicek 22
Size of subset (% of dataset)
0 10 20 30 40 50 60 70 80 90
Ano
maly
score
106
107
108
109
1010
1011
ANN w. 16 hiddenANN w. 32 hiddenANN w. 43 hidden
Iterations
1 2 3 4 5 6 7 8 9 10 11 12
Ra
tio
an
om
aly
sco
re o
f X
mo
d / X
va
l
0
1
2
3
4
5
6
7
8
9
ANN w. 43 hiddenANN w. 32 hiddenANN w. 16 hidden
ANN with 43 hidden nodes has 6-8 times higher anomaly score than validation set. 16 ~ factor 1.5
Integration options
Armin Wasicek 23
• Software Component (SWC) in AUTOSAR terminology
• Subscribe to relevant data via Virtual Function Bus (VFB)
• CAID integration options
o Standalone: on the CAN bus connecting the Central Gateway (CGW) to the OBD port
o Integrated: Part of the CGW
Related Work
• CAN message statistics [Hoppe et al., 2007]
• Entropy-based IDS [Muter et al., 2011]
• Commercial IDS/IPS: Deep Packet Inspection identifiesabnormal behavior
• Context-aware IDS [Wasicek and Weimerskirch, 2015]
24Armin Wasicek
Conclusion and Outlook
• Automotive systems are Cyber-Physical
• IDS need to recognize cyber and physical attacks
• Integrate with other security mechanisms
• Intelligently use the cloud to recognize attacks
• Faults, ageing, and repair effects are challenging
25Armin Wasicek
Response and Recovery
What to do after an intrusion has been detected?
• Not yet clear
• Depends on location, current state of vehicle
– Log threat
– Disable/inhibit features
– Service procedure
►What are means to react on intrusions/misuse?
Armin Wasicek 27
Automotive IDS Architecture
Armin Wasicek 28