Anomaly detection for vehicular networks for intrusion and malfunction detection

• US 9,792,435 B2
• Filed: 09/17/2015
• Issued: 10/17/2017
• Est. Priority Date: 12/30/2014
• Status: Active Grant

• Current Assignee: Battelle Memorial Institute
• Original Assignee: Battelle Memorial Institute
• Inventors: Harris, Brad, Sonalker, Anuja, Mayhew, Kevin
• Primary Examiner(s): To, Baotran N
• Assistant Examiner(s): Beheshti Shirazi, Sayed
• Application Number:
  US14/857,016
• Publication Number:
  US 20160188876A1
• Time in Patent Office: 761 Days
• CPC Class Codes: G06F 21/554 involving event detection and direct action
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]

Abstract

A security monitoring system for a Controller Area Network (CAN) comprises an Electronic Control Unit (ECU) operatively connected to the CAN bus. The ECU is programmed to classify a message read from the CAN bus as either normal or anomalous using an SVM-based classifier with a Radial Basis Function (RBF) kernel. The classifying includes computing a hyperplane curvature parameter γ of the RBF kernel as γ=ƒ(D) where ƒ( ) denotes a function and D denotes CAN bus message density as a function of time. In some such embodiments γ=ƒ(Var(D)) where Var(D) denotes the variance of the CAN bus message density as a function of time. The security monitoring system may be installed in a vehicle (e.g. automobile, truck, watercraft, aircraft) including a vehicle CAN bus, with the ECU operatively connected to the vehicle CAN bus to read messages communicated on the CAN bus. By not relying on any proprietary knowledge of arbitration IDs from manufacturers through their dbc files, this anomaly detector truly functions as a zero knowledge detector.

Claims

1. A Support Vector Machine (SVM) classifier training device comprising:

a computer including at least one microprocessor programmed to train a Support Vector Machine (SVM) one-class classifier using a Radial Basis Function (RBF) kernel K calculated using the equation;

K(x_i−

x_j)=e<sup>−

γ

∥

xi−

xj)∥

2</sup> where xε

and γ

>

0

where is a set of all real numbers, x_i and x_j are features of a training set, and γ

represents a curvature of a hyperplane and varies with message density D in time according to;

2. The SVM classifier training system of claim 1 wherein the operations further include:

determining a hyperplane granularity parameter ε

by repeating the training for a grid of candidate ε

values using a grid search.

3. The SVM classifier training system of claim 1 wherein the computer comprises a cloud-based computing resource, a parallel processor system, or a cluster computer.

4. The SVM classifier training system of claim 1 wherein γ

is calculated using the equation;

5. The SVM classifier training system of claim 1 wherein the data set contains only normal messages of a CAN bus and the training comprises training the SVM one-class classifier to minimize one of (i) false positives and (ii) false negatives.

6. The SVM classifier training system of claim 1 wherein the operations further comprise:

setting a tuning parameter ν

specifying a minimum number of Support Vectors to 520.

7. The SVM classifier training system of claim 1 wherein the computer is further programmed to:

receive at least one additional training subset; and

repeat the operations performed on the training set on the at least one additional training subset.

8. The SVM classifier training system of claim 1 wherein the training set comprises CAN bus messages acquired over at least one time interval from a vehicle, and the computer is further programmed to:

generate an SVM-based anomaly detector configured to execute on a vehicle-grade Electronic Control Unit (ECU) based on the trained SVM one-class classifier.

9. The SVM classifier training system of claim 8 wherein the operation of generating the SVM-based anomaly detector configured to execute on a vehicle-grade ECU comprises:

replacing the calculated high precision hyperplane curvature parameter γ

with a resource-constrained-hardware-compatible hyperplane curvature parameter γ

.

10. A system comprising:

a Support Vector Machine (SVM) classifier training device comprising a computer including at least one microprocessor programmed to train a Support Vector Machine (SVM) one-class classifier using a Radial Basis Function (RBF) kernel K calculated using the equation;

K(x_i−

x_j)=e<sup>−

γ

∥

xi−

xj)∥

2</sup> where xε

and γ

>

0

where is a set of all real numbers, x_i and x_j are features of a training set, and γ

represents a curvature of a hyperplane and varies with message density D in time according to;

11. The system of claim 10, further comprising:

an unseen ID unit implemented by the microprocessor configured to receive information from an observed ID list, wherein the alerting module is configured to receive information from the unseen ID unit and the SVM classifier and wherein;

the unseen ID unit receives a message from an electronic control unit (ECU) transceiver;

if the unseen ID unit determines that the message is not on an observed ID list, the unseen ID unit sends a binary indicator to the alerting module indicating that message is not on the observed ID list; and

if the unseen ID unit determines that the message is on the observed ID list, the unseen ID unit sends the message to the trained SVM one-class classifier.

12. A system as in claim 10 further comprising:

a vehicle including a vehicle controller area network (CAN) bus employing a message-based communication protocol; and

an electronic control unit (ECU) including an ECU microprocessor installed on the vehicle CAN bus and programmed to execute an anomaly detector comprising the trained SVM one-class classifier, the unseen ID unit, and the alerting module.