Image classification and accelerated classification training using deep learning image fingerprinting models and indexed embeddings
Image fingerprints (embeddings) are generated by an image fingerprinting model and indexed with an approximate nearest neighbors (ANN) model trained to identify the most similar fingerprint based on a subject embedding. For image matching, a score is provided that indicates a similarity between the input embedding and the most similar identified embedding, which allows for matching even when an image has been distorted, rotated, cropped, or otherwise modified. For image classification, the embeddings in the index are clustered and the clusters are labeled. Users can provide just a few images to add to the index as a labeled cluster. The ANN model returns a score and label of the most similar identified embedding for labeling the subject image if the score exceeds a threshold. As improvements are made to the image fingerprinting model, a converter model is trained to convert the original embeddings to be compatible with the new embeddings.
1. A computer-implemented method, comprising:
generating a first plurality of image fingerprint embeddings, the generating comprising:
providing each of a first plurality of images as input to an image fingerprinting model, wherein the image fingerprinting model is trained to generate an image fingerprint embedding comprising a multi-dimensional feature vector representing visual aspects of the input image, and
receiving each of the first plurality of image fingerprint embeddings as output from the image fingerprinting model;
receiving a user-defined label and a second plurality of images submitted by a user, wherein the images in the second plurality of images are distinct from the images in the first plurality of images;
generating a second plurality of image fingerprint embeddings, the generating comprising:
providing each of the second plurality of images as input to the image fingerprinting model, and
receiving each of the second plurality of image fingerprint embeddings as output from the image fingerprinting model;
training a user-defined image classifier based on and in response to receiving the user-defined label and the second plurality of images, the user-defined image classifier comprising an index and an approximate nearest neighbors model, the training comprising:
generating the index comprising the first plurality of image fingerprint embeddings and the second plurality of image fingerprint embeddings, wherein generating the index comprises:
generating at least one embedding cluster comprising at least a portion of the first plurality of image fingerprint embeddings;
generating a user-defined embedding cluster comprising the second plurality of image fingerprint embeddings;
applying a label to each of the at least one embedding cluster; and
applying the user-defined label to the user-defined embedding cluster, and
generating the approximate nearest neighbors model associated with the index, wherein:
the approximate nearest neighbors model is trained to classify subject images using the index; and
the training the user-defined image classifier does not include deep learning.
2. The computer-implemented method of claim 1 , wherein the second plurality of images comprises between twenty (20) and one hundred (100) images.
3. The computer-implemented method of claim 1 , further comprising analyzing a subject image, the analyzing comprising:
providing the subject image as input to the image fingerprinting model;
receiving a subject image fingerprint embedding as output from the image fingerprinting model;
providing the subject image fingerprint embedding as input to the user-defined image classifier trained to use the approximate nearest neighbors model and the index to classify the subject image, the classifying comprising retrieving a label of a most similar image fingerprint embedding from the index and generating a score indicating a similarity of the subject image fingerprint embedding to the most similar image fingerprint embedding;
analyzing the score received from the approximate nearest neighbors model; and
labeling the subject image based on the score.
4. The computer-implemented method of claim 3 , further comprising:
applying a security policy to the subject image based on the label.
5. The computer-implemented method of claim 3 , wherein:
analyzing the score comprises comparing the score to a threshold value;
labeling the subject image comprises labeling the subject image with the label of the most similar image fingerprint embedding when the score exceeds the threshold value; and
labeling the subject image comprises labeling the subject image with a default label when the score is below the threshold value.
6. The computer-implemented method of claim 3 , wherein the score comprises a value of zero to one hundred (0-100) that indicates an angular distance between the subject image fingerprint embedding and the most similar image fingerprint embedding.
7. The computer-implemented method of claim 3 , further comprising:
identifying the subject image based on a user performing an action including the subject image.
8. The computer-implemented method of claim 7 , wherein the action comprises one of:
uploading the subject image to a cloud application;
downloading the subject image from the cloud application;
deleting the subject image from the cloud application;
sharing the subject image on the cloud application;
moving the subject image within the cloud application; and
moving the subject image outside the cloud application.
9. The computer-implemented method of claim 1 , wherein the first plurality of image fingerprint embeddings comprises a plurality of negative image fingerprint embeddings, the method further comprising:
generating the plurality of negative image fingerprint embeddings, the generating comprising:
providing each of a plurality of negative images as input to the image fingerprinting model, and
receiving each of the plurality of negative image fingerprint embeddings as output from the image fingerprinting model; and
labeling each negative image fingerprint embedding with a default label in the index.
10. The computer-implemented method of claim 1 , further comprising:
providing a graphical user interface comprising user input elements selectable by a user for providing the second plurality of images and the user-defined label.
11. The computer-implemented method of claim 1 , further comprising:
receiving a third plurality of images selected by the user and a second user-defined label for the third plurality of images;
generating a third plurality of image fingerprint embeddings, the generating comprising:
providing each of the third plurality of images as input to the image fingerprinting model, and
receiving each of the third plurality of image fingerprint embeddings as output from the image fingerprinting model; and
training a second user-defined image classifier comprising a second index and a second approximate nearest neighbors model, the training comprising:
adding the third plurality of image fingerprint embeddings to the index to generate a second index and a second approximate nearest neighbors model associated with the second index,
clustering the third plurality of image fingerprint embeddings into a second user embedding cluster,
applying the second user-defined label to the second user embedding cluster, and
generating the second approximate nearest neighbors model associated with the second index.
12. The computer-implemented method of claim 11 , further comprising analyzing a subject image, the analyzing comprising:
providing the subject image as input to the image fingerprinting model;
receiving a subject image fingerprint embedding as output from the image fingerprinting model;
providing the subject image fingerprint embedding as input to the user-defined image classifier trained to use the approximate nearest neighbors model and the index to classify the subject image, the classifying comprising retrieving a first label of a most similar image fingerprint embedding from the index and generating a first score indicating a similarity of the subject image fingerprint embedding to the most similar image fingerprint embedding from the index;
providing the subject image fingerprint embedding as input to the second user-defined image classifier trained to use the second approximate nearest neighbors model and the second index to retrieve a second label of a most similar image fingerprint embedding from the second index and generate a second score indicating a similarity of the image subject fingerprint embedding to the most similar image fingerprint embedding from the second index;
analyzing the first score and the second score; and
labeling the subject image based on the first score and the second score.
13. The computer-implemented method of claim 12 , further comprising:
based on a determination that the first score is higher than the second score, labeling the subject image with the first label.
14. The computer-implemented method of claim 12 , further comprising:
based on a determination that the first score and the second score exceed a threshold value, labeling the subject image with the first label and the second label.
15. The computer-implemented method of claim 12 , further comprising:
based on a determination that the first score and the second score are below a threshold value, labeling the subject image with a default label.
16. A network security system, comprising:
one or more processors; and
one or more computer-readable memory devices, comprising:
an image fingerprinting model trained to generate an image fingerprint embedding comprising a multi-dimensional feature vector representing visual aspects of an input image,
a first plurality of image fingerprint embeddings generated by the image fingerprinting model from a first plurality of images, and
a classifier training component comprising instructions that, upon execution by the one or more processors, cause the one or more processors to:
receive a user-defined label and a second plurality of images uploaded by a user, wherein the images in the second plurality of images are distinct from the images in the first plurality of images;
generate a second plurality of image fingerprint embeddings, the instructions to generate comprising instructions that cause the one or more processors to:
provide each of the second plurality of images as input to the image fingerprint model, and
receive each of the second plurality of image fingerprint embeddings as output from the image fingerprinting model;
train a user-defined image classifier based on and in response to receiving the user-defined label and the second plurality of images, the user-defined image classifier comprising an index and an approximate nearest neighbors model, wherein the instructions to train comprise instructions to:
generate the index comprising the first plurality of image fingerprint embeddings and the second plurality of image fingerprint embeddings, wherein the instructions to generate the index comprises instructions to:
generate at least one embedding cluster comprising at least a portion of the first plurality of image fingerprint embeddings;
generate a user-defined embedding cluster comprising the second plurality of image fingerprint embeddings;
apply a label to each of the at least one embedding cluster; and
apply the user-defined label to the user-defined embedding cluster, and
generate the approximate nearest neighbors model associated with the index, wherein:
the approximate nearest neighbors model is trained to classify subject images using the index; and
the instructions to train does not include deep learning.
17. The network security system of claim 16 , the one or more computer-readable memory devices further comprising:
a subject image labeling component comprising instructions that, upon execution by the one or more processors, cause the one or more processors to:
provide a subject image as input to the image fingerprinting model;
receive a subject image fingerprint embedding as output from the image fingerprinting model;
provide the subject image fingerprint embedding as input to the user-defined image classifier trained to use the approximate nearest neighbors model and the index to classify the subject image, the classifying comprising retrieving a label of a most similar image fingerprint embedding from the index and generating a score indicating a similarity of the subject image fingerprint embedding to the most similar image fingerprint embedding;
analyze the score received from the approximate nearest neighbors model; and
label the subject image based on the score.
18. The network security system of claim 17 , the one or more computer-readable memory devices further comprising:
a data loss prevention component comprising instructions that, upon execution by the one or more processors, cause the one or more processors to:
apply a security policy to the subject image based on the label.
19. The network security system of claim 17 , the one or more computer-readable memory devices further comprising:
a plurality of user-defined image classifiers each comprising an index from a plurality of indexes and an approximate nearest neighbors model of a plurality of approximate nearest neighbors models associated with a respective one of the plurality of indexes generated by the classifier training component; and
wherein the subject image labeling component comprises instructions that cause the one or more processors to:
provide the subject image fingerprint embedding as input to each of the plurality of user-defined image classifiers, and
label the subject image based on scores from each of the plurality of user-defined image classifiers.
20. A computer-readable storage device having stored thereon instructions that, upon execution by one or more processors, cause the one or more processors to:
generate a first plurality of image fingerprint embeddings, the instructions to generate comprising instructions that cause the one or more processors to:
provide each of a first plurality of images as input to an image fingerprinting model, wherein the image fingerprinting model is trained to generate an image fingerprint embedding comprising a multi-dimensional feature vector representing visual aspects of the input image, and
receive each of the first plurality of image fingerprint embeddings as output from the image fingerprinting model;
receive a user-defined label and a second plurality of images uploaded by a user, wherein the images in the second plurality of images are distinct from the images in the first plurality of images;
generate a second plurality of image fingerprint embeddings, the instructions to generate comprising instructions that cause the one or more processors to:
provide each of the second plurality of images as input to the image fingerprinting model, and
receive each of the second plurality of image fingerprint embeddings as output from the image fingerprinting model;
train a user-defined image classifier based on and in response to receiving the user-defined label and the second plurality of images, the user-defined image classifier comprising an index and an approximate nearest neighbors model, the instructions to train comprising instructions to:
generate the index comprising the first plurality of image fingerprint embeddings and the second plurality of image fingerprint embeddings, wherein the instructions to generate the index comprises instructions to:
generate at least one embedding cluster comprising at least a portion of the first plurality of image fingerprint embeddings;
generate a user-defined embedding cluster comprising the second plurality of image fingerprint embeddings;
apply a label to each of the at least one embedding cluster; and
apply the user-defined label to the user-defined embedding cluster, and
generate the approximate nearest neighbors model associated with the index, wherein:
the approximate nearest neighbors model is trained to classify subject images using the index; and
the instructions to train the user-defined image classifier do not include deep learning.