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

Abstract

Methods for training a universal discriminator and an obfuscation network are provided. The methods include steps of: generating an obfuscated image by obfuscating an original image through the obfuscation network, characteristic information by applying learning operation to the obfuscated image through a surrogate network, a first discriminant score determining a degree of de-identification for the obfuscated image through the universal discriminator, and a second discriminant score determining whether the obfuscated image is real or fake through a regular discriminator, and thus training the obfuscation network through minimizing an accuracy loss of the surrogate network, and maximizing the first and second discriminant scores, wherein the universal discriminator has been trained by classifying de-identified images into positive or negative samples according to degrees of de-identification, generating discriminant scores determining degrees of de-identification for the samples through the universal discriminator, and training the universal discriminator by minimizing discriminator losses of the discriminant scores.