# US 12,223,734 B2 · Systems and methods for training machine-learned models with deviating intermediate representations

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

- Number: US 12,223,734 B2
- Document ID: US12223734B2
- Status: Granted
- Kind: B2
- Application number: 17/151,001
- Filed: 2021-01-15
- Published: 2021-09-09
- Granted: 2025-02-11
- Earliest claimed: 2020-03-05
- Assignees: Aurora Operations, Inc.
- Applicants: Aurora Operations, Inc.
- Inventors: Xuanyuan Tu; Raquel Urtasun; Tsun-Hsuan Wang; Sivabalan Manivasagam; Jingkang Wang; Mengye Ren
- Web page: https://patentgrep.com/patent/US12223734B2

## CPC

- G06V 20/56: exterior to a vehicle by using sensors mounted on the vehicle
- G01S 17/931: of land vehicles
- G05D 1/0088: {characterized by the autonomous decision making process, e.g. artificial intelligence, predefined behaviours (using knowledge based models G06N5/00)}
- G05D 1/0221: {involving a learning process}
- G05D 1/81: Handing over between on-board automatic and on-board manual control
- G06F 18/2163: {Partitioning the feature space}
- G06F 18/217: {Validation; Performance evaluation; Active pattern learning techniques}
- G06F 18/24: Classification techniques
- 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
- G06N 20/00: Machine learning
- G06V 10/764: using classification, e.g. of video objects
- G06V 10/82: using neural networks
- G01S 7/003: {Transmission of data between radar, sonar or lidar systems and remote stations}
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G08G 1/096791: {where the origin of the information is another vehicle}
- G08G 1/162: {event-triggered}
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G01S: RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G05D: SYSTEMS FOR CONTROLLING OR REGULATING NON-ELECTRIC VARIABLES
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G08G: TRAFFIC CONTROL SYSTEMS (guiding railway traffic, ensuring the safety of railway traffic B61L; arrangement of road signs or traffic signals E01F9/00; radar or analogous systems, sonar systems, lidar systems specially adapted for traffic control G01S13/91, G01S15/88, G01S17/88; {radar or analogous systems, sonar systems, lidar systems specially adapted for anti-collision purposes G01S13/93, G01S15/93, G01S17/93})

## Abstract

Systems and methods for vehicle-to-vehicle communications are provided. An adverse system can obtain sensor data representative of an environment proximate to a targeted system. The adverse system can generate an intermediate representation of the environment and a representation deviation for the intermediate representation. The representation deviation can be designed to disrupt a machine-learned model associated with the target system. The adverse system can communicate the intermediate representation modified by the representation deviation to the target system. The target system can train the machine-learned model associated with the target system to detect the modified intermediate representation. Detected modified intermediate representations can be discarded before disrupting the machine-learned model.

## Claims

### Claim 1

A computer-implemented method, the method comprising:
obtaining sensor data representative of a secondary environment proximate to an autonomous vehicle;
generating an intermediate representation for the autonomous vehicle based, at least in part, on the sensor data, wherein the intermediate representation is descriptive of at least a portion of the secondary environment;
determining an intermediate representation deviation for the intermediate representation based, at least in part, on the intermediate representation and a machine-learned model associated with the autonomous vehicle, by:
obtaining, via the machine-learned model, one or more ground truth bounding box proposals based, at least in part, on the intermediate representation,
obtaining, via the machine-learned model, one or more deviating bounding box proposals, and
determining the intermediate representation deviation for the intermediate representation based, at least in part, on a comparison between the one or more ground truth bound box proposals and the one or more deviating bounding box proposals;
generating data indicative of a modified intermediate representation based, at least in part, on the intermediate representation and the intermediate representation deviation; and
communicating the data indicative of the modified intermediate representation to a vehicle computing system associated with the autonomous vehicle.

### Claim 2 (depends on 1)

The computer-implemented method of claim 1, wherein the machine-learned model associated with the autonomous vehicle comprises a machine-learned model utilized by the vehicle computing system to detect one or more objects within a surrounding environment of the autonomous vehicle.

### Claim 3 (depends on 2)

The computer-implemented method of claim 2, wherein the machine-learned model is configured to output one or more bounding box proposals indicative of one or more objects within the surrounding environment of the autonomous vehicle based, at least in part, on one or more intermediate representations.

### Claim 4 (depends on 1)

The computer-implemented method of claim 1, wherein the intermediate representation deviation is based at least in part on an adversarial loss.

### Claim 5 (depends on 4)

The computer-implemented method of claim 4, wherein modifying the intermediate representation deviation for the intermediate representation based, at least in part, on the comparison between the one or more ground truth bound box proposals and the one or more deviating bounding box proposals comprise:
determining the adversarial loss for the intermediate representation deviation based, at least in part, on the one or more ground truth bound box proposals and the one or more deviating bounding box proposals; and
modifying the intermediate representation deviation based, at least in part, on the adversarial loss, wherein the intermediate representation deviation is modified to minimize the adversarial loss over the one or more deviating bounding box proposals.

### Claim 6 (depends on 5)

The computer-implemented method of claim 5, wherein each respective ground truth bounding box proposal of the one or more ground truth bounding box proposals comprises a respective ground truth class score indicative of respective ground truth object classification and one or more respective ground truth bounding box parameters indicative of a respective ground truth spatial location and one or more respective ground truth dimensions of the respective ground truth object classification, and
wherein each respective deviating bounding box proposal of the one or more respective deviating bounding box proposals comprises a respective deviating class score indicative of a respective deviating object classification and one or more respective deviating bounding box parameters indicative of a respective deviating spatial location and one or more respective deviating dimensions of the respective deviating object classification.

### Claim 7 (depends on 6)

The computer-implemented method of claim 6, wherein the adversarial loss is determined based, at least in part, on a difference between a ground truth class score corresponding to at least one ground truth bounding box proposal and a deviating class score corresponding to a deviating bounding box proposal corresponding to the at least one ground truth bounding box proposal.

### Claim 8 (depends on 6)

The computer-implemented method of claim 6, wherein the adversarial loss is determined based, at least in part, on a difference between one or more ground truth bounding box parameters corresponding to at least one ground truth bounding box proposal and one or more deviating bounding box parameters corresponding to a deviating bounding box proposal corresponding to the at least one ground truth bounding box proposal.

### Claim 9 (depends on 4)

The computer-implemented method of claim 4, wherein the computing system is onboard a transmitting autonomous vehicle physically located proximate to the autonomous vehicle, and wherein the intermediate representation deviation is associated with a first time.

### Claim 10 (depends on 9)

The computer-implemented method of claim 9, further comprising:
obtaining movement data indicative of a motion of the transmitting autonomous vehicle from the first time to a second time;
obtaining second sensor data representative of the secondary environment proximate to the autonomous vehicle at the second time;
generating a second intermediate representation for the autonomous vehicle based, at least in part, on the second sensor data; and
determining a second intermediate representation deviation for the second intermediate representation based, at least in part, on the intermediate representation deviation associated with the first time and the movement data.

### Claim 11

A computing system comprising:
one or more processors; and
one or more non-transitory computer-readable media that collectively store instructions that, when executed by the one or more processors, cause the system to perform operations, the operations comprising:
obtaining a plurality of intermediate representations associated with an autonomous vehicle, wherein each intermediate representation is descriptive of at least a portion of a secondary environment proximate to the autonomous vehicle at a plurality of times;
generating a surrogate machine-learned model based, at least in part, on the plurality of intermediate representations;
obtaining a target intermediate representation from the plurality of intermediate representations;
determining an intermediate representation deviation for the target intermediate representation based, at least in part, on the target intermediate representation and the surrogate machine-learned model by:
obtaining, via the surrogate machine-learned model, one or more ground truth bounding box proposals based, at least in part, on the target intermediate representation,
obtaining, via the surrogate machine-learned model, one or more deviating bounding box proposals, and
determining the intermediate representation deviation for the target intermediate representation based, at least in part, on a comparison between the one or more ground truth bound box proposals and the one or more deviating bounding box proposals;
generating data indicative of a modified intermediate representation based, at least in part, on the target intermediate representation and the intermediate representation deviation; and
communicating the data indicative of the modified intermediate representation to a vehicle computing system associated with the autonomous vehicle.

### Claim 12 (depends on 11)

The computing system of claim 11, wherein each intermediate representation of the plurality of intermediate representations is generated by a first portion of a machine-learned model associated with the autonomous vehicle, and wherein a first portion of the surrogate machine-learned model is trained to output a surrogate intermediate representation, wherein the surrogate intermediate representation comprises one or more similarities to the plurality of intermediate representations.

### Claim 13 (depends on 11)

The computing system of claim 11, wherein generating the surrogate machine-learned model based, at least in part, on the plurality of intermediate representations comprises:
obtaining sensor data representative of surrogate environment proximate to the autonomous vehicle; and
generating the surrogate machine-learned model based, at least in part, on the plurality of intermediate representations and the sensor data.

### Claim 14 (depends on 13)

The computing system of claim 13, wherein generating the surrogate machine-learned model based, at least in part, on the plurality of intermediate representations further comprises:
generating, via a first portion of the surrogate machine-learned model, a surrogate intermediate representation based, at least in part, on the sensor data;
generating, via a machine-learned discriminator model, a discriminator loss based, at least in part, on the surrogate intermediate representation and at least one of the plurality of intermediate representations; and
training the surrogate machine-learned model to minimize the discriminator loss.

### Claim 15 (depends on 14)

The computing system of claim 14, wherein the discriminator loss is indicative of a difference between the surrogate intermediate representation and the at least one intermediate representation.

### Claim 16 (depends on 11)

The computing system of claim 11, wherein a second portion of the surrogate machine-learned model is configured to output one or more bounding box proposals indicative of one or more objects within the secondary environment proximate to the autonomous vehicle based, at least in part, on the target intermediate representation.

### Claim 17 (depends on 16)

The computing system of claim 16, wherein the intermediate representation deviation is based at least in part on an adversarial loss.

### Claim 18

An autonomous vehicle comprising:
one or more sensors;
one or more processors; and
one or more tangible, non-transitory, computer readable media that collectively store instructions that when executed by the one or more processors cause the one or more processors to perform operations, the operations comprising:
obtaining, via the one or more sensors, sensor data representative of a surrounding environment of the autonomous vehicle;
generating, via a first portion of a machine-learned model, an intermediate representation based, at least in part, on the sensor data, wherein the intermediate representation is descriptive of at least a portion of the surrounding environment of the autonomous vehicle;
determining an intermediate representation deviation for the intermediate representation based, at least in part, on the intermediate representation and the machine-learned model, by:
obtaining, via the machine-learned model, one or more ground truth bounding box proposals based, at least in part, on the intermediate representation,
obtaining, via the machine-learned model, one or more deviating bounding box proposals, and
determining the intermediate representation deviation for the intermediate representation based, at least in part, on a comparison between the one or more ground truth bound box proposals and the one or more deviating bounding box proposals;
generating, data indicative of a modified intermediate representation based, at least in part, on the intermediate representation and the intermediate representation deviation; and
communicating the data indicative of the modified intermediate representation to one or more devices associated with a target autonomous vehicle.

### Claim 19 (depends on 18)

The autonomous vehicle of claim 18, wherein the target autonomous vehicle is configured to utilize a second portion of the machine-learned model to detect one or more objects within a surrounding environment of the target autonomous vehicle.

### Claim 20 (depends on 18)

The autonomous vehicle of claim 18, wherein the machine-learned model is trained to detect the modified intermediate representation.