System and method for image de-identification
AssigneeDe-Identification Ltd.
InventorsGil Perry, Sella Blondheim, Eliran Kuta, Yoav Hacohen
Life of this patent
Estimating the term…
Abstract
System and method for training a human perception predictor to determine level of perceived similarity between data samples, the method including: receiving at least one media file, determining at least one identification region for each media file, applying at least one transformation on each identification region for each media file until at least one modified media file is created, receiving input regarding similarity between each modified media file and the corresponding received media file, and training a machine learning model with an objective function configured to predict similarity between media files by a human observer in accordance with the received input.
Claims12
A method for de-identification processing of image files whilst maintaining a level of similarity between an original and a modified image file, the method comprising:
- receiving at least one image file;
- determining, for the at least one image file, at least one face for de-identification processing;
- applying at least one transformation on the at least one face to create at least one modified image file, thereby limiting identification by computerized algorithms;
- and
- validating the at least one modified image file by:
- receiving, from a computerized human vision predictor trained to recognize similarity between image files similarly to the perception of an actual human, a measure of human-perceptual similarity between the at least one modified image file and the corresponding received image file;
- receiving a recognition score from an existing computerized face recognition algorithm configured to identify the at least one face in the at least one modified image file;
- determining if the measure of human-perceptual similarity exceeds a visual similarity threshold and if the recognition score is below a de-identification threshold;
- and
- iteratively updating the at least one applied transformation and applying such updated transformation to the received at least one image file if either the recognition score is not below the de-identification threshold or the measure of human-perceptual similarity does not exceed the visual similarity threshold, else outputting the modified image file.
A method of synthesizing media files, the method comprising:
- receiving a media file;
- determining at least one identification region for the received media file;
- modifying the at least one identification region;
- applying at least one transformation on the modified identification region;
- generating a synthetic media file by reconstruction of the transformed identifications region;
- repeating applying and generating until the applied transformation causes reduction of confidence value for the corresponding received media file by at least one classifier, wherein the generated synthetic media file and the received media file are recognized as including similar faces by a simulator of human perception;
- training a parametric model on a dataset of identification regions to synthesize a synthesized media file given a set of parameters;
- obtaining the parameters representing the identification region as depicted in the received media file;
- and
- modifying the parameters to represent a modified version of the identification region so as to retain resemblance to the identification region of the received media file while reducing the recognition score of the at least one classifier below a predefined threshold.
Dependencies are inferred from the claim text.