Automated detection of tampered images

1. A method for training a fake image detector, the method comprising:

training a plurality of autoencoders to generate fake images, each autoencoder including a first decoder and a second decoder, the training including:

training the first decoder to generate images incorporating features of a first person into images of a second person; and

training the second decoder to generate images incorporating features of the second person into images of the first person;

using the trained plurality of autoencoders to generate a training set of images, the training set including, for each autoencoder:

positive examples generated by the first decoder from input images depicting the second person and the second decoder from input images depicting the first person, wherein the positive examples include the second person depicted with features of the first person and the first person depicted with features of the second person, and

negative examples generated by the first decoder from input images depicting the first person and the second decoder from input images depicting the second person; and

training an image fake model to identify fake images using the training set of images wherein a fake image depicts a computer-generated composite of features from two or more persons.

2. The method of claim 1 , wherein training the image fake model to identify fake images using the training set of images includes:

providing, as inputs to the image fake model, the positive examples and the negative examples generated by the first decoder and the second decoder, the positive example images labeled as fake and the negative example images labeled as real;

generating, by the image fake model, predictions of whether the inputs are fake or real; and

comparing the predictions to the labels to determine whether the image fake model is trained.

3. The method of claim 2 , wherein a degree of difference between the predictions and the labels is used in a cost function to further train the image fake model.

4. The method of claim 1 , wherein the images incorporating features of the first person into images of the second person are images of the first person's face on the second person's body.

5. The method of claim 1 , wherein training the plurality of autoencoders to generate fake images further includes overfitting the first decoder to leave input images depicting the first person substantially unchanged and overfitting the second decoder to leave input images depicting the second person substantially unchanged.

6. The method of claim 1 , wherein the image fake model is a neural network, a support vector machine model, a gradient boosted decision tree, or a linear regression model.

7. A fake image detection system stored on a computer-readable medium, wherein the fake image detection system is manufactured by a process comprising:

training a plurality of autoencoders to generate fake images, each autoencoder including a first decoder and a second decoder, the training including:

training the first decoder to generate images incorporating features of a first person into images of a second person; and

training the second decoder to generate images incorporating features of the second person into images of the first person;

using the trained plurality of autoencoders to generate a training set of images, the training set including, for each autoencoder:

positive examples generated by the first decoder from input images depicting the second person and the second decoder from input images depicting the first person, wherein the positive examples include the second person depicted with features of the first person and the first person depicted with features of the second person, and

negative examples generated by the first decoder from input images depicting the first person and the second decoder from input images depicting the second person; and

training an image fake model to identify fake images using the training set of images wherein a fake image depicts a computer-generated composite of features from two or more persons.

8. The fake image detection system of claim 7 , wherein training the image fake model to identify fake images using the training set of images includes:

providing, as inputs to the image fake model, the positive examples and the negative examples generated by the first decoder and the second decoder, the positive example images labeled as fake and the negative example images labeled as real;

generating, by the image fake model, predictions whether the inputs are fake or real; and

comparing the predictions to the labels to determine whether the image fake model is trained.

9. The fake image detection system of claim 7 , wherein the images incorporating features of the first person into images of the second person are images of the first person's face on the second person's body.

10. The fake image detection system of claim 7 , wherein training the plurality of autoencoders to generate fake images further includes overfitting the first decoder to leave input images depicting the first person substantially unchanged and overfitting the second decoder to leave input images depicting the second person substantially unchanged.

11. A method for detecting one or more fake images, the method comprising:

receiving an image;

providing the image as input to a trained image fake model, wherein the trained image fake model was trained by a process comprising:

training a plurality of autoencoders to generate fake images, each autoencoder including a first decoder and a second decoder, the training including:

training the first decoder to generate images incorporating features of a first person into images of a second person; and

training the second decoder to generate images incorporating features of the second person into images of the first person;

using the trained plurality of autoencoders to generate a set of training images, the set of training images including, for each autoencoder:

positive examples generated by the first decoder from input images depicting the second person and the second decoder from input images depicting the first person, wherein the positive examples include the second person depicted with features of the first person and the first person depicted with features of the second person, and

negative examples generated by the first decoder from input images depicting the first person and the second decoder from input images depicting the second person; and

training the image fake model to identify a fake image using the set of training images, wherein a fake image depicts a computer-generated composite of features from two or more persons; and

generating, by the image fake model, an output indicating whether the image is fake.

12. The method of claim 11 , further comprising:

associating metadata with the image indicating whether the image is fake based on the output of the image fake model.

13. The method of claim 12 , wherein the output from the image fake model includes a probability that the image is fake, and the metadata is added to the image if the probability exceeds a threshold.

14. The method of claim 11 , wherein the process by which the trained image fake model was trained further comprised:

using the trained plurality of autoencoders to generate the set of training images, the set of training images further including, for each autoencoder, negative examples generated by the first decoder from input images depicting the first person and the second decoder from input images depicting the second person.

15. The method of claim 11 , wherein the images incorporating features of the first person into images of the second person are images of the first person's face on the second person's body.

16. The method of claim 11 , wherein the process by which the trained image fake model was trained further comprised:

overfitting the first decoder to leave input images depicting the first person substantially unchanged and overfitting the second decoder to leave input images depicting the second person substantially unchanged.

17. The method of claim 11 , wherein the images generated by the first and second decoders both include artifacts.

18. The method of claim 11 , wherein the image fake model is a neural network, a support vector machine model, a gradient boosted decision tree, or a linear regression model.