Adversarial attack detection and avoidance in computer vision
Techniques for adversarial attack avoidance for machine learning (ML) are disclosed. These techniques include receiving one or more images at a trained ML model and receiving attack data at the ML model. The techniques further include predicting an object depicted in the one or more images using the ML model, based on the one or more images, metadata relating to the one or more images, and the attack data. The ML model uses the metadata to prevent the attack data from changing a result of the predicting.
1. A method, comprising:
receiving, from a computing system implementing a machine learning (ML) model, one or more images depicting an object, wherein the ML model is trained using training data representing the object and additional training metadata relating to the object, the additional training metadata comprising temporal metadata that includes two or more training images depicting the object at different times;
receiving, from the computing system, attack data at the ML model; and
outputting, to the computing system, an object classification of the object depicted in the one or more images predicted using the ML model, based on the one or more images, the additional training metadata, and the attack data;
wherein the ML model uses the additional training metadata to prevent the attack data from changing the object classification of the object.
2. The method of claim 1 , wherein the attack data comprises adversarial attack data intended to change the object classification of the object by the ML model.
3. The method of claim 1 , wherein the ML model is trained using a multi-dimensional array relating to the training data and the additional training metadata.
4. The method of claim 1 , wherein the additional training metadata further comprises object relationship data for the object.
5. The method of claim 4 , wherein the object relationship data comprises a graph relating to the object depicted in the one or more images.
6. The method of claim 5 , wherein the graph comprises a plurality of vertices and edges relating to sub-components of the object.
7. The method of claim 6 , wherein the graph is generated using a second ML model to identify the sub-components.
8. The method of claim 7 , wherein the graph is further generated by identifying bounding boxes for each of the sub-components and calculating a respective centroid for each bounding box, each of the vertices in the graph relating to one or more of the centroids.
9. A non-transitory computer-readable medium including computer program code that, when executed by operation of one or more computer processors, performs operations comprising:
receiving, from a computing system implementing a machine learning (ML) model, one or more images depicting an object wherein the ML model is trained using training data representing the object and additional training metadata relating to the object, the additional training metadata comprising temporal metadata that includes two or more training images depicting the object at different times;
receiving, from the computing system, attack data at the ML model; and
outputting, to the computing system, an object classification of the object depicted in the one or more images predicted using the ML model, based on the one or more images, the additional training metadata, and the attack data;
wherein the ML model uses the additional training metadata to prevent the attack data from changing the object classification of the object.
10. The non-transitory computer-readable medium of claim 1 , wherein the additional training metadata further comprises object relationship data for the object.
11. The non-transitory computer-readable medium of claim 3 , wherein the object relationship data comprises a graph relating to the object depicted in the one or more images.
12. The non-transitory computer-readable medium of claim 9 , wherein the attack data comprises adversarial attack data intended to change the object classification of the object by the ML model.
13. The non-transitory computer-readable medium of claim 9 , wherein the ML model is trained using a multi-dimensional array relating to the training data and the additional training metadata.
14. The non-transitory computer-readable medium of claim 11 , wherein the graph comprises a plurality of vertices and edges relating to sub-components of the object.
15. A system, comprising:
a computer processor; and
a memory having instructions stored thereon which, when executed on the computer processor, performs operations comprising:
receiving, from the system, one or more images depicting an object at a machine learning (ML) model implemented by the system, wherein the ML model is trained using training data representing the object and additional training metadata relating to the object, the additional training metadata comprising temporal metadata that includes two or more training images depicting the object at different times;
receiving, from the system, attack data at the ML model; and
outputting, to the system, an object classification of the object depicted in the one or more images predicted using the ML model, based on the one or more images, the additional training metadata, and the attack data;
wherein the ML model uses the additional training metadata to prevent the attack data from changing the object classification of the object.
16. The system of claim 14 , wherein the additional training metadata further comprises object relationship data for the object.
17. The system of claim 7 , wherein the object relationship data comprises a graph relating to the object depicted in the one or more images.
18. The system of claim 15 , wherein the attack data comprises adversarial attack data intended to change the object classification of the object by the ML model.
19. The system of claim 15 , wherein the ML model is trained using a multi-dimensional array relating to the training data and the additional training metadata.
20. The system of claim 17 , wherein the graph comprises a plurality of vertices and edges relating to sub-components of the object.