Spatio-temporal anomaly detection in computer networks using graph convolutional recurrent neural networks (GCRNNs)
In one embodiment, a device receives sensor data from a plurality of nodes in a computer network. The device uses the sensor data and a graph that represents a topology of the nodes in the network as input to a graph convolutional neural network. The device provides an output of the graph convolutional neural network as input to a convolutional long short-term memory recurrent neural network. The device detects an anomaly in the computer network by comparing a reconstruction error associated with an output of the convolutional long short-term memory recurrent neural network to a defined threshold. The device initiates a mitigation action in the computer network for the detected anomaly.
1. A method comprising:
receiving, at a device, sensor data from a plurality of nodes in a computer network;
using, by the device, the sensor data and a graph that represents a topology of the nodes in the network as input to a graph convolutional neural network;
providing, by the device, an output of the graph convolutional neural network as input to a convolutional long short-term memory recurrent neural network, wherein the graph convolutional neural network is configured to produce its output based in part on a spatial dependency between the nodes in the computer network;
detecting, by the device, an anomaly in the computer network by comparing a reconstruction error associated with an output of the convolutional long short-term memory recurrent neural network to a defined threshold, wherein the convolutional long short-term memory recurrent neural network is configured to produce its output based on a temporal dependency of the sensor data; and
initiating, by the device, a mitigation action in the computer network for the detected anomaly.
2. The method as in claim 1 , wherein the mitigation action comprises one of: blocking or dropping traffic in the computer network, removing one or more of the nodes associated with the detected anomaly from the routing topology, or sending an alert regarding the detected anomaly to a user interface.
3. The method as in claim 1 , wherein the sensor data comprises log data from the plurality of nodes in the computer network regarding operational characteristics of the nodes.
4. The method as in claim 1 , further comprising:
receiving, at the device, the defined threshold from a user interface.
5. The method as in claim 1 , further comprising:
training the graph convolutional neural network and convolutional long short-term memory recurrent neural network using back propagation.
6. The method as in claim 5 , wherein the training uses training data comprising time series of the sensor data indicative of normal operations of the computer network.
7. An apparatus, comprising:
one or more network interfaces to communicate with a computer network;
a processor coupled to the network interfaces and configured to execute one or more processes; and
a memory configured to store a process executable by the processor, the process when executed configured to:
receive sensor data from a plurality of nodes in a computer network;
use the sensor data and a graph that represents a topology of the nodes in the network as input to a graph convolutional neural network;
provide an output of the graph convolutional neural network as input to a convolutional long short-term memory recurrent neural network, wherein the graph convolutional neural network is configured to produce its output based in part on a spatial dependency between the nodes in the computer network;
detect an anomaly in the computer network by comparing a reconstruction error associated with an output of the convolutional long short-term memory recurrent neural network to a defined threshold, wherein the convolutional long short-term memory recurrent neural network is configured to produce its output based on a temporal dependency of the sensor data; and
initiate a mitigation action in the computer network for the detected anomaly.
8. The apparatus as in claim 7 , wherein the mitigation action comprises one of: blocking or dropping traffic in the computer network, removing one or more of the nodes associated with the detected anomaly from the routing topology, or sending an alert regarding the detected anomaly to a user interface.
9. The apparatus as in claim 7 , wherein the sensor data comprises log data from the plurality of nodes in the computer network regarding operational characteristics of the nodes.
10. The apparatus as in claim 7 , wherein the process when executed is further configured to:
receive the defined threshold from a user interface.
11. The apparatus as in claim 7 , wherein the process when executed is further configured to:
train the graph convolutional neural network and convolutional long short-term memory recurrent neural network using back propagation.
12. The apparatus as in claim 11 , wherein the training uses training data comprising time series of the sensor data indicative of normal operations of the computer network.
13. A tangible, non-transitory, computer-readable medium storing program instructions that cause a processor of a device to execute a process comprising:
receiving, at the device, sensor data from a plurality of nodes in a computer network;
using, by the device, the sensor data and a graph that represents a topology of the nodes in the network as input to a graph convolutional neural network;
providing, by the device, an output of the graph convolutional neural network as input to a convolutional long short-term memory recurrent neural network, wherein the graph convolutional neural network is configured to produce its output based in part on a spatial dependency between the nodes in the computer network;
detecting, by the device, an anomaly in the computer network by comparing a reconstruction error associated with an output of the convolutional long short-term memory recurrent neural network to a defined threshold, wherein the convolutional long short-term memory recurrent neural network is configured to produce its output based on a temporal dependency of the sensor data; and
initiating, by the device, a mitigation action in the computer network for the detected anomaly.
14. The computer-readable medium as in claim 13 , wherein the mitigation action comprises one of: blocking or dropping traffic in the computer network, removing one or more of the nodes associated with the detected anomaly from the routing topology, or sending an alert regarding the detected anomaly to a user interface.
15. The computer-readable medium as in claim 13 , wherein the sensor data comprises log data from the plurality of nodes in the computer network regarding operational characteristics of the nodes.
16. The computer-readable medium as in claim 13 , wherein the process further comprises:
receiving, at the device, the defined threshold from a user interface.
17. The computer-readable medium as in claim 13 , wherein the process further comprises:
training the graph convolutional neural network and convolutional long short-term memory recurrent neural network using back propagation.