Data set generation for performance overhead prediction of obfuscated code

Software obfuscation is widely used as a protection mechanism against reverse engineering attacks, slowing down the activity of crackers and hiding the actual logic of the program code. This protection is generally used in commercial products with the most sales during the early days after the release. However, there are some disadvantages to using this security technique, and the most serious one is indeed a drop in performance. Effects of protections on performance can be challenging to predict a priori and can slow down the work of developers, who waste their time searching the optimal combination of obfuscation techniques for their programs. This work aims to solve this problem and find a correlation between a specific obfuscation technique, the program that needs to be obfuscated, and the resulting performance overhead. To achieve this result, we needed to follow multiple steps. First of all, we have selected a test set of applications suitable for our purposes. Then we have extracted some features to characterize every program and see how the performance evaluations vary when comparing the executions of the program with their obfuscated versions. The previous steps generated a data set that can be used as input to build and train some machine learning models that will be used to predict these performance faults.