# US 2021/0067549 A1 · Anomaly detection with graph adversarial training in computer systems

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0067549 A1
- Document ID: US20210067549A1
- Status: Published application
- Kind: A1
- Application number: 17/004,752
- Filed: 2020-08-27
- Published: 2021-03-04
- Earliest claimed: 2019-08-29
- Assignees: NEC Laboratories America, Inc.
- Applicants: NEC Laboratories America, Inc.
- Inventors: Zhengzhang Chen; Jiaping Gui; Haifeng Chen; Junghwan Rhee; Shen Wang
- Web page: https://patentgrep.com/patent/US20210067549A1

## CPC

- H04L 63/1466: {Active attacks involving interception, injection, modification, spoofing of data unit addresses, e.g. hijacking, packet injection or TCP sequence number attacks}
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/08: Learning methods
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/09: Supervised learning
- G06N 3/094: Adversarial learning
- H04L 63/1425: {Traffic logging, e.g. anomaly detection}
- H04L 63/20: {for managing network security; network security policies in general (filtering policies H04L63/0227)}
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Methods and systems for detecting and responding to an intrusion in a computer network include generating an adversarial training data set that includes original samples and adversarial samples, by perturbing one or more of the original samples with an integrated gradient attack to generate the adversarial samples. The original and adversarial samples are encoded to generate respective original and adversarial graph representations, based on node neighborhood aggregation. A graph-based neural network is trained to detect anomalous activity in a computer network, using the adversarial training data set. A security action is performed responsive to the detected anomalous activity.

## Claims

### Claim 1

A method for detecting and responding to an intrusion in a computer network, comprising:
generating an adversarial training data set that includes original samples and adversarial samples, by perturbing one or more of the original samples with an integrated gradient attack to generate the adversarial samples;
encoding the original and adversarial samples to generate respective original and adversarial graph representations, based on node neighborhood aggregation;
training a graph-based neural network to detect anomalous activity in a computer network, using the adversarial training data set; and
performing a security action responsive to the detected anomalous activity.

### Claim 2 (depends on 1)

The method of claim 1, wherein the original samples include graph information.

### Claim 3 (depends on 2)

The method of claim 2, wherein each adversarial sample includes a copy of one of the original samples, with an edge added or removed.

### Claim 4 (depends on 1)

The method of claim 1, wherein encoding the original and adversarial samples includes a plurality of stages, with each stage including a neighborhood aggregation and a representation update.

### Claim 5 (depends on 4)

The method of claim 4, wherein the neighborhood aggregation includes a mean aggregation.

### Claim 6 (depends on 4)

The method of claim 4, wherein the representation update for each stage is based on each output from all previous stages.

### Claim 7 (depends on 4)

The method of claim 4, wherein encoding the original and adversarial samples further includes a bottleneck that limits a dimensionality of a final graph representation.

### Claim 8 (depends on 1)

The method of claim 1, wherein perturbing one or more of the original samples includes determining an integrated gradient for a perturbation to determine how much the perturbation will affect a prediction outcome for an unlabeled node.

### Claim 9 (depends on 10)

The method of claim 10, wherein the security action is selected from the group consisting of shutting down devices, stopping or restricting a type of network communication, enabling or disabling a connection between two devices, raising an alert to a system administrator, and changing a security policy level.

### Claim 10

A system for detecting and responding to an intrusion in a computer network, comprising:
a hardware processor;
a memory;
a graph-based neural network model, configured to detect anomalous activity in a computer network;
a model trainer, stored in the memory and executed by the hardware processor, configured to generate an adversarial training data set that includes original samples and adversarial samples, by perturbing one or more of the original samples with an integrated gradient attack to generate the adversarial samples; to encode the original and adversarial samples to generate respective original and adversarial graph representations, based on node neighborhood aggregation; and to train the graph-based neural network to detect anomalous activity in a computer network, using the adversarial training data set; and
a security console configured to perform a security action responsive to the detected anomalous activity.

### Claim 11 (depends on 10)

The system of claim 10, wherein the original samples include graph information.

### Claim 12 (depends on 11)

The system of claim 11, wherein each adversarial sample includes a copy of one of the original samples, with an edge added or removed.

### Claim 13 (depends on 10)

The system of claim 10, wherein the model trainer is further configured to encode the original and adversarial samples in a plurality of stages, with each stage including a neighborhood aggregation and a representation update.

### Claim 14 (depends on 13)

The system of claim 13, wherein the neighborhood aggregation includes a mean aggregation.

### Claim 15 (depends on 13)

The system of claim 13, wherein the representation update for each stage is based on each output from all previous stages.

### Claim 16 (depends on 13)

The system of claim 13, wherein the model trainer is further configured to use a bottleneck that limits a dimensionality of a final graph representation.

### Claim 17 (depends on 10)

The system of claim 10, wherein the model trainer is further configured to determine an integrated gradient for a perturbation to determine how much the perturbation will affect a prediction outcome for an unlabeled node.

### Claim 18 (depends on 10)

The system of claim 10, wherein the security action is selected from the group consisting of shutting down devices, stopping or restricting a type of network communication, enabling or disabling a connection between two devices, raising an alert to a system administrator, and changing a security policy level.
