Random walk in hidden messages
Starting these posts to help track the articles I like …
I was browsing the archive and found this very recent submission on algorithms called StegaStamp.
StegaStamp is all about steganography which is the practice of concealing a data within another type of data. This data can be images, music, text, or entire files.
The authors imagine a time when we move between the everyday and the internet seamlessly. For example, we walk through a museum and the little signs next to the artwork will send us to Wikipedia for further information. Sounds kind of awesome but Google Glass never did catch on.
The authors design a deep learning model to hide and recognize hyperlinks in images. An autoencoder takes an image and hyperlink to output an encoded image. One can then take a picture of a printed version of the encoded image and then decode the hyperlink. The encoder and decoder networks are trained on corruptions caused by typical printing.
The authors point to another interesting paper from Li Fei-Fei’s lab that is worth a read. HiDDeN takes advantage image manipulations like compression and cropping noise to hide data in images.
Though steganography is not a new concept, both papers are a reminder that people are terribly clever to take advantage of flaws to hide and share information.
