System and method for image de-identification to humans while remaining recognizable by machines
A system and method for de-identification of an image may include receiving an input image of a human face; iteratively modifying the input image to produce an output image of the human face until: (a) a threshold indicating how unlikely a human is to associate the input image with the output image is reached, and (b) a threshold indicating an ability of a computerized unit to associate the input image with the output image is reached; and providing the output image.
1. A method of image de-identification, the method comprising:
receiving an input image of a human face;
iteratively modifying the input image to produce an output image of the human face by modifying style parameters while maintaining structure features of the input image until:
a score indicating how unlikely a human is to associate the input image with the output image reaches a first threshold indicating that the human would not be able to associate the input image with the output image, and
a score indicating an ability of a computerized unit to associate the input image with the output image reaches a second threshold indicating that the computerized unit would be able to associate the input image with the output image; and
providing the output image,
wherein modifying the input image is performed by a deep neural network, wherein the style parameters comprise high level features of the neural network, and wherein the neural network modifies the input image by modifying the high level features while maintaining the structure features,
wherein modifying the style parameters while maintaining the structure features is carried out with application of a loss function :
=λ 0 ∥Θ 0 ( X src )−Θ 0 ( X D )∥−λ 1 ∥Θ 1 ( X src )−Θ 1 ( X D )∥ . . . −λ k ∥Θ k ( X src )−Θ k ( X D )∥
where k is a scale index, λ is a weight, Θ is a parameter of the neural network, X src is the input image, X D is the output image, and Θ k are the high level features.
2. The method of claim 1 , comprising optimizing the de-identification by:
calculating, in at least some of the iterations: (a) a first score value produced by a first function reflecting the likelihood that a human would associate the face in the input image with the face in the output image; and (b) a second score value produced by a second function reflecting the ability of a computerized unit to associate the face in the input image with the face in the output image; and
providing the output image if a function of the first and second score values reaches a threshold.
3. The method of claim 2 , comprising, providing the output image if a sequence of values produced by the function of the first and second score values converges to a value.
4. The method of claim 1 , comprising, providing the modified image upon determining a threshold number of iterations was reached.
5. A method of generating a neural network (NN) model, the method comprising:
training a NN including the model by, for each input image in a set of input images of faces:
generating, using the model, an output image of the face;
calculating, by a first function, a dis-similarity score reflecting the likelihood that a human would associate the face in the input image with the face in the output image;
calculating, by a second function, an identity-similarity score reflecting an ability of a computerized unit to associate the face in the input image with the face in the output image; and
modifying parameters of the model using a loss function designed to manipulate style features while maintaining structure features such that the dis-similarity score reaches a first threshold indicating that the human would not be able to associate the face in the input image with the face in the output image, and the identity-similarity score reaches a second threshold indicating that the computerized unit would be able to associate the face in the input image with the face in the output image,
wherein the style features comprise high level features of the model,
wherein modifying the style features while maintaining the structure features is carried out with application of a loss function :
=λ 0 ∥Θ 0 ( X src )−Θ 0 ( X D )∥−λ 1 ∥Θ 1 ( X src )−Θ 1 ( X D )∥ . . . −λ k ∥Θ k ( X src )−Θ k ( X D )∥
where k is a scale index, λ is a weight, Θ is a parameter of the neural network, X src is the input image, X D is the output image, and Θ k are the high level features.
6. The method of claim 5 , comprising, modifying the parameters of the model such that a sequence of values produced by a function of the first and second scores converges to a value.
7. The method of claim 5 , comprising, terminating the training when determining at least one of:
one or more scores produced by the first function reached a first threshold,
one or more scores produced by the second function reached a second threshold, and
a threshold number of processed images was reached.
8. The method of claim 5 comprising, modifying the parameters of the model based on an aggregated score of a set of input images.
9. The method of claim 5 , comprising terminating the training when determining at least one of: scores produced by the first function and scores produced by the second function have reached one of: a minimal value and a maximal value.
10. The method of claim 5 , comprising terminating the training when determining a function of at least one of: scores produced by the first function and scores produced by the second function has reached one of: a minimal value and a maximal value.
11. A system of image de-identification, the system comprising:
a memory; and
a processor configured to:
receive an input image of a human face;
iteratively modify the input image to produce an output image of the human face by modifying style parameters while maintaining structure features of the input image until:
a score indicating how unlikely a human is to associate the input image with the output image reaches a first threshold indicating that the human would not be able to associate the input image with the output image, and
a score indicating an ability of a computerized unit to associate the input image with the output image reaches a second threshold indicating that the computerized unit would be able to associate the input image with the output image; and
providing the output image,
wherein the processor is further configured to modify the input image by a deep neural network, wherein the style parameters comprise high level features of the neural network, and wherein the neural network modifies the input image by modifying the high level features while maintaining the structure features,
wherein modifying the style parameters while maintaining the structure features is carried out with application of a loss function :
=λ 0 ∥Θ 0 ( X src )−Θ 0 ( X D )∥−λ 1 ∥Θ 1 ( X src )−Θ 1 ( X D )∥ . . . −λ k ∥Θ k ( X src )−Θ k ( X D )∥
where k is a scale index, λ is a weight, Θ is a parameter of the neural network, X src is the input image, X D is the output image, and Θ k are the high level features.
12. The system of claim 11 , wherein the processor is configured to optimize the de-identification by:
calculating, in at least some of the iterations: (a) a first score value produced by a first function reflecting the likelihood that a human would associate the face in the input image with the face in the output image; and (b) a second score value produced by a second function reflecting the ability of a computerized unit to associate the face in the input image with the face in the output image; and
providing the output image if a function of the first and second score values reaches a threshold.
13. The system of claim 12 , wherein the processor is configured to provide the output image if a sequence of values produced by the function of the first and second score values converges to a value.
14. The system of claim 11 , wherein the processor is configured to provide the modified image upon determining a threshold number of iterations was reached.
15. The method of claim 1 , wherein the style parameters comprise high level features of a facial recognition model.
16. The method of claim 1 , wherein iteratively modifying comprises using the output image of a first iteration is as the input image to a subsequent iteration.
17. The method of claim 5 , wherein the style parameters comprise high level features of the NN.
18. The system of claim 11 , wherein the processor is configured to modify the input image by a deep neural network, wherein the style parameters comprise high level features of the neural network, and wherein the neural network modifies the input image by modifying the high level features while maintaining the structure features.