Towards the prediction of performance degradation of obfuscated code

Software obfuscation is the most common protection against Man-at-the-End attacks. However, it can be challenging to evaluate the effectiveness of each technique and the effects it has on the resulting software performance. This thesis addresses the problem of performance prediction in software obfuscation, focusing on obfuscation techniques applied to C source code. The final goal of this work is to create a dataset of accurate information about the execution of various applications matched with the respective performance variations when different obfuscation techniques are applied. To this purpose, an environment has been created to run tests and automatically collect reliable data about software execution and its performance. It includes a solution to run and trace applications with inputs generated with a fuzzer and a testbed to reliably measure the performance drop due to the applied obfuscation techniques. The created dataset has also been used to train an initial set of traditional machine learning models to test if predicting performance variation is feasible, with promising results.