Demystify scrambled and encoded images with auto-encoders

Monitoring systems have become more and more popular. A typical example is a smart home monitoring system which provides peace of mind for people staying far away from their homes. This system can be deployed for different purposes, such as to take care elderly staying at home alone or to provide anti-theft when nobody is at home. In this type of monitoring systems, there is always a problem with protecting the privacy and security of captured images. This problem is often addressed by either scrambling or encoding techniques that add noise in or to distort the images. In the IEEE Big Data 2022 Cup, the challenge is to detect and perform matching between an original image and its scrambled or encoded version. Generally, if we are able to train a machine learning model that would correctly identify the matching, it will be proof that the applied scrambling or encoding technique is not strong enough to protect data and hence it needs to be strengthened further. In this paper, we present an auto-encoder architecture that allows us to demystify scrambled or encoded images as long as the scrambling or encoding techniques are deterministic. Source code is publicly available
¹
.