System and method for performing facial image anonymization

1. A method of performing facial image anonymization, the method comprising:

obtaining, by at least one processor of a first computing device, a first image data element, depicting a first face;

using, by the at least one processor, a first neural network (NN), to encode the first image data element into a first latent vector;

using, by the at least one processor, a transformation module, wherein the transformation module comprises at least one second NN, trained to receive the first latent vector and produce therefrom an anonymized latent vector; and

transmitting, by the at least one processor, the anonymized latent vector, via a communication network to a remote second computing device,

wherein the second computing device is adapted to decode the anonymized latent vector into a second image data element, depicting an anonymized face,

wherein training the at least one second NN comprises:

introducing the first image data element as a first input to a face-recognition (FR) module;

introducing the second image data element as a second input to the FR module;

introducing the first latent vector to at least one feature classifier as a first input;

introducing the anonymized latent vector to the at least one feature classifier as a second input; and

using output of the FR module and output of the at least one feature classifier as supervisory data, such that the first image data element and second image data element comprise the same features according to the at least one feature classifier, and do not pertain to the same identity according to the FR module.

2. The method of claim 1 , wherein the anonymized latent vector comprises at least one first group of parameters, representing features pertaining to identity, and at least one second group of parameters, representing features that do not pertain to identity.

3. The method of claim 2 wherein the transformation module is adapted to produce the anonymized latent vector by modifying a value of at least one parameter of the at least one first group, and not modifying a value of at least one parameter of the at least one second group.

4. The method of claim 1 , wherein the anonymized face is not identified by a face recognition algorithm as portraying the same person as the first face.

5. The method of claim 1 , wherein the output of the FR module comprises a probability that a face depicted in the first image data element and a face depicted in in the second image data element pertain to the same person and wherein output of the at least one feature classifier comprises a probability that at least one feature of the first image data element is equal to the same feature in the second image data element.

6. A method of performing facial image anonymization, the method comprising:

obtaining, by a processor of a first computing device, a first digital image of a face;

encoding, by the processor, the first digital image into a first latent vector;

modifying, by the processor, the first latent vector, using at least one second NN, trained to receive the first latent vector and to produce therefrom an anonymized latent vector; and

transmitting, by the processor, the anonymized latent vector to a second computing device,

wherein the second computing device is adapted to decode the anonymized latent vector into a second digital image, depicting an anonymized face,

wherein training the at least one second NN comprises:

introducing the first digital image as a first input to a face-recognition (FR) module;

introducing the second digital image element as a second input to the FR module;

introducing the first latent vector to at least one feature classifier as a first input;

introducing the anonymized latent vector to the at least one feature classifier as a second input; and

using output of the FR module and output of the at least one feature classifier as supervisory data, such that the first digital image and second digital image comprise the same features according to the at least one feature classifier, and do not pertain to the same identity according to the FR module.

7. The method of claim 6 , wherein the output of the FR module comprises a probability that a face depicted in the first image data element and a face depicted in in the second image data element pertain to the same person and wherein output of the at least one feature classifier comprises a probability that at least one feature of the first image data element is equal to the same feature in the second image data element.

8. A system for performing anonymization of a facial image, the system comprising: a non-transitory memory device, wherein modules of instruction code are stored, and a processor associated with the memory device, and configured to execute the modules of instruction code, whereupon execution of said modules of instruction code, the processor is further configured to:

obtain a first image data element, depicting a first face;

use a first neural network (NN) to encode the first image data element into a first latent vector;

modify the first latent vector, using at least one second NN, trained to receive the first latent vector and to produce therefrom an anonymized latent vector; and

transmit the anonymized latent vector via a communication network to a second, remote computing device, adapted to decode the anonymized latent vector into a second image data element, depicting an anonymized face,

wherein training the at least one second NN comprises:

introducing the first digital image as a first input to a face-recognition (FR) module;

introducing the second digital image element as a second input to the FR module;

introducing the first latent vector to at least one feature classifier as a first input;

introducing the anonymized latent vector to the at least one feature classifier as a second input; and

using output of the FR module and output of the at least one feature classifier as supervisory data, such that the first digital image and second digital image comprise the same features according to the at least one feature classifier, and do not pertain to the same identity according to the FR module.

9. The system of claim 8 ,

wherein the anonymized latent vector comprises a first group of parameters representing features pertaining to identity, and a second group of parameters representing features that do not pertain to identity, and

modifying the second latent vector.

10. The system of claim 9 wherein the processor is configured produce the anonymized latent vector by modifying a value of at least one parameter of the first group, and not modifying a value of at least one parameter of the second group.

11. The system of claim 8 , wherein the anonymized face is not identified by a face recognition algorithm as portraying the same person as the first face.

12. The system of claim 8 , wherein the output of the FR module comprises a probability that a face depicted in the first image data element and a face depicted in in the second image data element pertain to the same person and wherein output of the at least one feature classifier comprises a probability that at least one feature of the first image data element is equal to the same feature in the second image data element.