Federated graph neural network for fast anomaly detection in controller area networks

Inventors

Zhang, Hengrun • Zeng, Kai

Assignees

George Mason University

Interested in licensing this patent?

MTEC can help explore whether this patent might be available for licensing for your application.

Publication Number

US-12488049-B2

Patent

Publication Date

2025-12-02

Expiration Date


Abstract

A method and apparatus for intrusion detection includes generating graph data from a sequence of messages in a communication network. The graph data denotes a node for each message identifier in the sequence of messages and an edge for each pair of consecutive message identifiers. Pair counts are generated for edges in the graph data as a number of times the associated pair of consecutive message identifiers occurs in the sequence. Input feature vectors are generated for nodes in the graph data. The vectors include data content of messages associated with the node and a pair count for each edge connected to the node. The input feature vectors are processed through a first graph neural network, based on the graph data, and then through one or more layers of a classifier to classify the sequence of messages as containing an anomaly or not.

Core Innovation

The invention provides a computer-implemented intrusion detection system that generates graph data from a sequence of messages in a communication network, where each message includes data content and a message identifier. The graph data denotes a node for each message identifier in the sequence of messages, and an edge for each pair of consecutive message identifiers, with the edge linking two nodes. For edges, a pair count is generated as a number of times the associated pair of consecutive message identifiers occurs in the sequence of messages, and for nodes, an input feature vector is generated including the data content of messages and a pair count for each edge connected to the node.

The invention processes the input feature vectors through a first graph neural network based on the graph data to produce first output feature vectors. It classifies the sequence of messages as containing an anomaly or not by processing the first output feature vectors through one or more first output layers. In an apparatus implementation, one or more computers and a non-transitory computer-readable medium are configured to perform the same graph-data generation, feature-vector generation, first graph neural network processing, and anomaly/not-anomaly classification.

The invention further supports a two-stage classifier approach in which, when an anomaly is involved, an additional path processes input feature vectors using a second graph neural network to produce second output feature vectors and classifies an anomaly using one or more second output layers. The document describes a second-stage anomaly classifier architecture that includes one or more output layers such as an openmax layer, and supports classifying known attack types while handling unknown anomalies.

Claims Coverage

The document includes three independent claims. Each independent claim centers on generating graph data from a sequence of messages using message identifiers and consecutive identifier pairs, forming node input feature vectors from message data content plus edge pair counts, and using a first graph neural network with one or more output layers to classify whether the message sequence contains an anomaly or not.

Consecutive-identifier graph data and edge pair counts from message sequences

Generating graph data from a sequence of messages in a communication network, the graph data denoting a node for each message identifier and an edge for each pair of consecutive message identifiers, and generating for edges a pair count as a number of times the associated pair of consecutive message identifiers occurs in the sequence of messages.

Node input feature vectors from message data content and incident edge pair counts

Generating, for nodes in the graph data, an input feature vector including data content of messages that include the message identifier associated with the node and a pair count for each edge connected to the node.

First graph neural network processing and anomaly-or-not classification via first output layers

Processing the input feature vectors through a first graph neural network, based on the graph data, to produce first output feature vectors, and classifying the sequence of messages as containing an anomaly or not by processing the first output feature vectors through one or more first output layers.

Conditional second graph neural network anomaly classification via second output layers

Classifying the sequence of messages as containing an anomaly or not including, when the sequence contains an anomaly, processing the input feature vectors using a second graph neural network to generate second output feature vectors and classifying the anomaly by processing those second output feature vectors through one or more second output layers.

Communication-network-coupled message graph generation and processing

In the apparatus, receiving a sequence of messages from a communication network and causing one or more computers to generate the graph data, generate node input feature vectors, process them using a first graph neural network, and classify whether the sequence contains an anomaly or not using one or more first output layers.

Controller Area Network (CAN) bus implementation as the communication network

In the apparatus, implementing the communication network as a Controller Area Network (CAN) bus.

Across the independent claims, the core coverage is the conversion of message sequences into directed graph data using message identifiers and consecutive identifier pairs, augmenting node representations with message data content plus incident edge pair counts, and using a first graph neural network with one or more output layers to classify anomaly versus not-anomaly. Dependent refinements within the independent-claim families add an alternate conditional path using a second graph neural network and second output layers for anomaly classification, and specify that the communication network may be a Controller Area Network (CAN) bus.

Stated Advantages

Fast detection.

Improved performance described for federated learning under non-IID vehicle state data.

Classifying known attack types while handling unknown anomalies.

Preserving privacy.

Documented Applications

Attacks including injection, suspension, replay, spoofing, and fuzzy attacks.

Driving scenarios and vehicle states.

JOIN OUR MAILING LIST

Stay Connected with MTEC

Keep up with active and upcoming solicitations, MTEC news and other valuable information.