Learning to process images depicting faces without leveraging sensitive attributes in deep learning models

1. A system, comprising:

a processor circuit; and

a memory storing instructions which when executed by the processor circuit cause the processor circuit to:

generate, by a neural network of a variational autoencoder, a latent vector for a first input image, a plurality of weights of the neural network optimized during training to not consider a sensitive attribute in images, the neural network trained based on: (i) a master model trained to detect the sensitive attribute in images, and (ii) a reconstruction loss in generating reconstructed training images, the reconstruction loss based on a difference between a training image and the reconstructed training image;

generate, by the neural network, a first reconstructed image by sampling the latent vector for the first input image; and

perform an image processing operation on the first reconstructed image.

2. The system of claim 1 , the image processing operation comprising one or more of: (i) determining whether a face is depicted in the first reconstructed image, and (ii) determining whether the face depicted in the first reconstructed image is expressing one or more emotions.

3. The system of claim 1 , wherein the training images comprise a plurality of training images, the memory storing instructions which when executed by the processor circuit cause the processor circuit to:

receive a first training image of the plurality of training images;

generate, by the neural network of the variational autoencoder, a first reconstructed training image by sampling the latent vector for the first training image;

determine a reconstruction loss incurred in generating the first reconstructed training image based at least in part on: (i) a difference of the first training image and the first reconstructed training image, and (ii) the master model;

determine a total loss based at least in part on the reconstruction loss and a classification loss; and

optimize a plurality of weights of the neural network of the variational autoencoder based on a backpropagation operation and the determined total loss, the optimization to train the neural network to not consider the sensitive attribute when processing the first reconstructed image.

4. The system of claim 3 , the memory storing instructions to determine the classification loss for the first reconstructed training image which when executed by the processor circuit cause the processor circuit to:

compute a difference of a predetermined latent vector for the first training image and an output of a convolutional neural network receiving the first reconstructed training image as input; and

compute a square of the computed difference.

5. The system of claim 4 , the memory storing instructions which when executed by the processor circuit cause the processor circuit to:

determine the total loss for the first reconstructed training image based on a sum of the classification loss and the reconstruction loss.

6. The system of claim 5 , the memory storing instructions to determine the reconstruction loss which when executed by the processor circuit cause the processor circuit to:

compute a first difference between the first training image and the first reconstructed training image;

compute a square of the first difference;

compute a second difference between a first output of the master model and a second output of the master model, the first output based on the first training image, the second output based on the first reconstructed training image;

compute a square of the second difference; and

compute a difference of the square of the first difference and the square of the second difference.

7. The system of claim 3 , wherein the master model is of a plurality of master models, wherein the sensitive attribute is of a plurality of sensitive attributes, wherein the plurality of sensitive attributes comprise: (i) a gender, (ii) a race, (iii) an age, (iv) a skin tone, and (v) a religion of a person depicted in images, each master model trained not to consider one of the sensitive attributes when processing the first reconstructed image.

8. A non-transitory computer-readable storage medium storing instructions that when executed by a processor cause the processor to:

generate, by a neural network of a variational autoencoder, a latent vector for a first input image, a plurality of weights of the neural network optimized during training to not consider a sensitive attribute in images, the neural network trained based on: (i) a master model trained to detect the sensitive attribute in images, and (ii) a reconstruction loss in generating reconstructed training images, the reconstruction loss based on a difference between a training image and the reconstructed training image;

generate, by the neural network, a first reconstructed image by sampling the latent vector for the first input image; and

perform an image processing operation on the first reconstructed image.

9. The non-transitory computer-readable storage medium of claim 8 , the image processing operation comprising one or more of: (i) determining whether a face is depicted in the first reconstructed image, and (ii) determining whether the face depicted in the first reconstructed image is expressing one or more emotions.

10. The non-transitory computer-readable storage medium of claim 8 , wherein the training images comprise a plurality of training images, further storing instructions that when executed by the processor cause the processor to:

receive a first training image of the plurality of training images;

generate, by the neural network of the variational autoencoder, a first reconstructed training image by sampling the latent vector for the first training image;

determine a reconstruction loss incurred in generating the first reconstructed training image based at least in part on: (i) a difference of the first training image and the first reconstructed training image, and (ii) the master model;

determine a total loss based at least in part on the reconstruction loss and a classification loss; and

optimize a plurality of weights of the neural network of the variational autoencoder based on a backpropagation operation and the determined total loss, the optimization to train the neural network to not consider the sensitive attribute when processing the first reconstructed image.

11. The non-transitory computer-readable storage medium of claim 10 , storing instructions to determine the classification loss for the first reconstructed training image that when executed by the processor cause the processor to:

compute a difference of a predetermined latent vector for the first training image and an output of a convolutional neural network receiving the first reconstructed training image as input; and

compute a square of the computed difference.

12. The non-transitory computer-readable storage medium of claim 11 , storing instructions that when executed by the processor cause the processor to:

determine the total loss for the first reconstructed training image based on a sum of the classification loss and the reconstruction loss.

13. The non-transitory computer-readable storage medium of claim 12 , storing instructions to determine the reconstruction loss which when executed by the processor cause the processor to:

compute a first difference between the first training image and the first reconstructed training image;

compute a square of the first difference;

compute a second difference between a first output of the master model and a second output of the master model, the first output based on the first training image, the second output based on the first reconstructed training image;

compute a square of the second difference; and

compute a difference of the square of the first difference and the square of the second difference.

14. The non-transitory computer-readable storage medium of claim 10 , wherein the master model is of a plurality of master models, wherein the sensitive attribute is of a plurality of sensitive attributes, wherein the plurality of sensitive attributes comprise: (i) a gender, (ii) a race, (iii) an age, (iv) a skin tone, and (v) a religion of a person depicted in images, each master model trained not to consider one of the sensitive attributes when processing images.

15. A method, comprising:

generating, by a neural network of a variational autoencoder, a latent vector for a first input image, a plurality of weights of the neural network optimized during training to not consider a sensitive attribute in images, the neural network trained based on: (i) a master model trained to detect the sensitive attribute in images, and (ii) a reconstruction loss in generating reconstructed training images, the reconstruction loss based on a difference between a training image and the reconstructed training image;

generating, by the neural network, a first reconstructed image by sampling the latent vector for the first input image; and

performing an image processing operation on the first reconstructed image.

16. The method of claim 15 , the image processing operation comprising one or more of: (i) determining whether a face is depicted in the first reconstructed image, and (ii) determining whether the face depicted in the first reconstructed image is expressing one or more emotions.

17. The method of claim 16 , wherein the training images comprise a plurality of training images, the method further comprising:

receiving a first training image of the plurality of training images;

generating, by the neural network of the variational autoencoder, a first reconstructed training image by sampling the latent vector for the first training image;

determining a reconstruction loss incurred in generating the first reconstructed training image based at least in part on: (i) a difference of the first training image and the first reconstructed training image, and (ii) the master model;

determining a total loss based at least in part on the reconstruction loss and a classification loss; and

optimizing a plurality of weights of the neural network of the variational autoencoder based on a backpropagation operation and the determined total loss, the optimization to train the neural network to not consider the sensitive attribute when processing the first reconstructed image.

18. The method of claim 17 , further comprising determining the classification loss for the first reconstructed training image by:

computing a difference of a predetermined latent vector for the first training image and an output of a convolutional neural network receiving the first reconstructed training image as input; and

computing a square of the computed difference.

19. The method of claim 18 , further comprising:

determining the total loss for the first reconstructed training image based on a sum of the classification loss and the reconstruction loss.

20. The method of claim 19 , wherein the master model is of a plurality of master models, wherein the sensitive attribute is of a plurality of sensitive attributes, wherein the plurality of sensitive attributes comprise: (i) a gender, (ii) a race, (iii) an age, (iv) a skin tone, and (v) a religion of a person depicted in images, each, master model trained not to consider one of the sensitive attributes when processing images, the method further comprising:

computing a first difference between the first training image and the first reconstructed training image;

computing a square of the first difference;

computing a second difference between a first output of the master model and a second output of the master model, the first output based on the first training image, the second output based on the first reconstructed training image;

computing a square of the second difference; and

computing a difference of the square of the first difference and the square of the second difference.