US Patent 11544460 Adversarial anonymization and preservation of content

Systems and methods for anonymizing content suggestive of a particular characteristic while preserving relevant content are disclosed. An example method may be performed by one or more processors of a protection system and include defining an anonymization loss indicative of an accuracy at which a trained discriminator model can predict a particular characteristic, defining a content loss indicative of a difference between latent representations of versions of a document, defining a combined objective function incorporating the anonymization and content losses, extracting and anonymizing suggestive content from training documents while preserving relevant content, and adversarially training, using the associated accuracies and differences in the combined objective function, a transformation model to transform a given document representative of credentials of a given person possessing the particular characteristic into an anonymized document maximizing a predicted uncertainty of the trained discriminator model while simultaneously maximizing an amount of relevant information about the person preserved.