Understanding image-set classifi ers for future evaluation of adversarial pro les to gain control of our own privacy.

Abstract

An abundance of images can be found on social media platforms nowadays. These images, uploaded by their users, can give us sensitive insights and information about the person behind it using machine learning techniques. In this work, we propose a framework aiding us in investigating the reaction of different set-based image classi fiers when controlling two aspects of picture sets, its dimensions and distribution. The extensive framework allows custom creation of user pro les according to rules, pretrained models on eleven of MS COCO's super-categories and two different implementations of set-based image classifi ers. The ultimate goal is to understand the workings of such methods that can conceivably be used by malicious actors wanting to infer privacy-sensitive information from pictures. That way we can deduce useful information from these fi ndings helping future research to craft adversarial techniques to help minimize privacy infringement.