Methods for training universal discriminator capable of determining degrees of de-identification for images and obfuscation network capable of obfuscating images and training devices using the same

A method for training an obfuscation network capable of obfuscating images, comprising steps of:

• (a) in response to acquiring a first original image, a first learning device performing or supporting another device to perform processes of inputting the first original image to an obfuscation network, and allowing the obfuscation network to obfuscate the first original image to thereby generate a first obfuscated image corresponding to the first original image;
• (b) the first learning device performing or supporting another device to perform processes of (i) (i-1) inputting the first obfuscated image into a surrogate network having at least one or more trained parameters, and allowing the surrogate network to apply at least one learning operation to the first obfuscated image by using the trained parameters and thus to generate characteristic information corresponding to the first obfuscated image, and (i-2) generating an accuracy loss by referring to the characteristic information and its corresponding ground truth or by referring to a specific task output and its corresponding ground truth, wherein the task specific output is generated by using the characteristic information, (ii) inputting the first obfuscated image to a universal discriminator, having been trained by a second learning device such that the universal discriminator generates discriminant scores corresponding to images inputted thereto according to each degree of de-identification for each of the images, and allowing the universal discriminator to generate a first discriminant score corresponding to the first obfuscated image according to a degree of the de-identification for the first obfuscated image, and (iii) inputting the first obfuscated image to a regular discriminator and allowing the regular discriminator to generate a second discriminant score corresponding to the first obfuscated image by determining whether the first obfuscated image is real or fake;
• and
• (c) the first learning device performing or supporting another device to perform processes of training the obfuscation network such that the accuracy loss is minimized and such that the first discriminant score and the second discriminant score are maximized.

A learning device for training a universal discriminator capable of determining each degree of de-identification for each image, comprising:

• at least one memory that stores instructions;
• and
• at least one processor configured to execute the instructions to perform or support another device to perform:
• (I) processes of (i) transmitting de-identified images to at least one labeler, wherein the de-identified images are generated by de-identifying original images, and (ii) allowing the labeler to classify the de-identified images into positive samples or negative samples according to each degree of the de-identification for each of the de-identified images, wherein a positive reference value is labelled as ground truth for the positive samples and a negative reference value is labelled as ground truth for the negative samples, (II) processes of inputting each of the positive samples and each of the negative samples to the universal discriminator, and allowing the universal discriminator to generate each of discriminant scores corresponding to each of the positive samples and each of the negative samples according to each degree of the de-identification for each of the positive samples and each of the negative samples, and (III) processes of generating discriminator losses by comparing the discriminant scores for each of the positive samples and each of the negative samples with their corresponding ground truths, and thus training the universal discriminator in a direction of minimizing the discriminator losses.

A first learning device for training an obfuscation network capable of obfuscating images, comprising:

• at least one memory that stores instructions;
• and
• at least one processor configured to execute the instructions to perform or support another device to perform:
• (I) in response to acquiring a first original image, processes of inputting the first original image to an obfuscation network, and allowing the obfuscation network to obfuscate the first original image to thereby generate a first obfuscated image corresponding to the first original image, (II) processes of (i) (i-1) inputting the first obfuscated image into a surrogate network having at least one or more trained parameters, and allowing the surrogate network to apply at least one learning operation to the first obfuscated image by using the trained parameters and thus to generate characteristic information corresponding to the first obfuscated image, and (i-2) generating an accuracy loss by referring to the characteristic information and its corresponding ground truth or by referring to a specific task output and its corresponding ground truth, wherein the task specific output is generated by using the characteristic information, (ii) inputting the first obfuscated image to a universal discriminator, having been trained by a second learning device such that the universal discriminator generates discriminant scores corresponding to images inputted thereto according to each degree of de-identification for each of the images, and allowing the universal discriminator to generate a first discriminant score corresponding to the first obfuscated image according to a degree of the de-identification for the first obfuscated image, and (iii) inputting the first obfuscated image to a regular discriminator and allowing the regular discriminator to generate a second discriminant score corresponding to the first obfuscated image by determining whether the first obfuscated image is real or fake, and (III) processes of training the obfuscation network such that the accuracy loss is minimized and such that the first discriminant score and the second discriminant score are maximized.