Unredactor VS DeepMalwareDetector

Interestingly, some years back (perhaps 12-15 years?) someone developed a program that would examine the font a physically redacted document was written in, and the spacing to try to unredact it, with some relatively decent success as only a set combination of words/letters etc. could fill a specific redacted portion. Of course the larger the redacted block, the harder it becomes. It was interesting none the less, not sure what happened to it though. This: https://github.com/gt0410/Unredactor is similar, but not what I was thinking of, and this: https://hackaday.com/2008/08/01/exposing-poorly-redacted-pdfs/ may also prove interesting for you.

What brought me to this research was these studies [1] [2], which demonstrates how image-based malware classification can be done using a CNN (convolutional neural network). Since I had a bit of a background with malware, and I recently completed a CNN model, I figured I would try to do something similar. It was only after investigating different materials I hit a bit of a roadblock. I found this one dataset, malimg [3], which is made up of PE files that have been converted into images already. I didn't want to just use the images, I wanted to demonstrate how to get them, only the method used to classify them turned out to be a bit out of my depth, kind of like this whole project, it's discussed in Section 4.2 of this paper [4] . There's also this set [5], which contains the pixel content for each file record. And as for the static disassembly you mention, I think you are right, the training data might not exist. During my investigation the best I could find was this study [6].