Systems and methods for privacy-enabled biometric processing
In one embodiment, a set of feature vectors can be derived from any biometric data, and then using a deep neural network (“DNN”) on those one-way homomorphic encryptions (i.e., each biometrics' feature vector) can determine matches or execute searches on encrypted data. Each biometrics' feature vector can then be stored and/or used in conjunction with respective classifications, for use in subsequent comparisons without fear of compromising the original biometric data. In various embodiments, the original biometric data is discarded responsive to generating the encrypted values. In another embodiment, the homomorphic encryption enables computations and comparisons on cypher text without decryption. This improves security over conventional approaches. Searching biometrics in the clear on any system, represents a significant security vulnerability. In various examples described herein, only the one-way encrypted biometric data is available on a given device. Various embodiments restrict execution to occur on encrypted biometrics for any matching or searching.
1. A privacy-enabled identification system comprising:
at least one processor operatively connected to a memory;
a classification component executed by the at least one processor, including a classification model including one or more deep neural networks (“DNNs”), the one or more DNNs trained on distance measurable homomorphic encrypted feature vector and respective label inputs, and wherein the DNN is configured to:
accept as an input distance measurable homomorphic encrypted feature vectors, the distance measurable homomorphic encrypted feature vectors generated as a one way encoding of plain text identification information of a first identification data type for an entity input to at least one first pre-trained neural network; and
predict a match to a label for identification or return unknown responsive to input of at least one distance measurable homomorphic encrypted feature vector produced by the at least one first neural network from identification information of the first identification data type.
2. The system of claim 1 , wherein the one or more DNNs are configured to identify an entity based on predicting a match to the label for identification information of a first identification data type that meets a threshold probability.
3. The system of claim 1 , further comprising a generation component executed by the at least one processor, including an embedding model, wherein the embedding model includes the at least one first pre-trained neural network, wherein the at least one first pre-trained neural network is configured to generate the one way encoding of plain text identification information responsive to input of the plain text identification information to the first pre-trained neural network.
4. The system of claim 3 , wherein the generation component further comprises:
a plurality of pre-trained embedding models associated with respective data types executed by a plurality of pre-trained neural networks, wherein each embedding model output is used to train a paired classification model executed by one or more respective DNNs, and
wherein the one or more respective DNNs are trained to classify the distance measurable homomorphic encrypted feature vectors generated by paired pre-trained neural networks based on the respective data types.
5. The system of claim 4 , wherein each pairing of pre-trained embedding model and classification model are configured to predict the match to the label for identification based on input of a new identification input to respective pre-trained neural networks and using output of a distance measurable homomorphic encrypted feature vector of the new identification input to the one or more respective DNNs.
6. The system of claim 5 , wherein the generation component further comprises multiple embedding models implemented by multiple neural networks for processing respective types of identification information.
7. The system of claim 1 , wherein the classification component further comprises multiple DNNs, each DNN associated with an identification information data type and a respective pre-trained neural network.
8. The system of claim 7 , wherein at least one respective pair of a DNN and the embedding model are configured to identify an entity based on a plain text input of audio identification information to the embedding model.
9. The system of claim 7 , wherein at least one DNN is configured to identify an entity based on input of distance measurable homomorphic encrypted feature vector produced from the audio identification information.
10. The system of claim 1 , wherein the classification component is configured with a plurality of modes of execution, including an enrollment mode configured to generate or accept a label for identification and associated encrypted feature vectors for linking the label to an entity.
11. The system of claim 1 , wherein the embedding model and respective pre-trained neural networks are pre-trained prior to enrollment of an entity to be identified or authenticated.
12. A computer implemented method for privacy-enabled biometric authentication, the method comprising:
instantiating, by at least one processor, a classification network including a classification model, the classification model including a deep neural network (“DNN”);
accepting, by the DNN, as an input distance measurable homomorphic encrypted feature vectors, wherein the distance measurable homomorphic encrypted feature vectors are generated as a one way encoding of plain text identification information input to at least one first neural network;
training, by at least one processor, the DNN on distance measurable homomorphic encrypted feature vectors and respective label inputs; and
predicting, by the DNN, a match to a label for identification or returning unknown responsive to input of at least one distance measurable homomorphic encrypted feature vector produced by the at least one first neural network.
13. The method of claim 12 , wherein the method further comprising confirming identification of the entity based on predicting a match to the label for identification information of a first identification data type that meets a threshold probability.
14. The method of claim 12 , further comprising instantiating a generation component, including an embedding model, wherein the embedding model includes the at least one first pre-trained neural network, and wherein the at least one first pre-trained neural network is configured to generate the one way encoding of plain text identification information responsive to input of the plain text identification information to the first pre-trained neural network.
15. The method of claim 14 , wherein the method further comprises generating, by the first pre-trained neural network the one way encoding of plain text identification information.
16. The method of claim 15 , wherein the method further comprises:
instantiating, by the at least one processor, a plurality embedding models including respective ones of a plurality of pre-trained neural networks, each embedding model paired with respective classification models executed by respective DNNs, and
wherein the respective DNNs of are trained to classify the one way encodings produced by the plurality of pre-trained neural networks of respective embedding models.
17. The method of claim 16 , wherein the method further comprises predicting the match to the label for identification based on input of a new identification input to respective pre-trained neural networks and using output of a distance measurable homomorphic encrypted feature vector of the new identification input to respective DNNs.
18. The method of claim 14 , wherein instantiating the generation component includes instantiating multiple embedding models implemented by multiple pre-trained neural networks for processing respective types of identification information.
19. The method of claim 12 , wherein instantiating the classification network further comprises instantiating multiple DNNs, each DNN associated with an identification information data type and a respective pre-trained neural network.
20. The method of claim 19 , wherein predicting a match to a label includes identifying an entity based on a plain text input of audio identification information to an embedding model to produce a distance measurable encrypted feature vectors input into a classification model trained on the distance measurable encrypted feature vectors of the audio identification information.
21. The method of claim 12 , wherein the method further comprises triggering one of a plurality of modes of execution, including an enrollment mode configured to accept a label for identification and associated encrypted feature vectors for an entity.