Estimating similarity between differently compiled procedures using neural networks

Dual licensing is a form of licensing where the same software product can be used under two different licenses. One license is usually free but introduces major restrictions on the usage of the software product, while the other allows a greater degree of freedom in using it, but at a specified monetary price. Although this form of licensing has a lot of benefits, one of its drawbacks is the increased possibility of license violation and copyright infringement. This paper deals with the problem of detecting a certain procedure from the source code in a set of disassembled binary procedures. The problem is made more difficult by not knowing which compiler was used to produce the binaries. The paper proposes an approach that falls into the category of Siamese neural networks and uses a combination of LSTM and DNN architectures. The proposed approach is evaluated independently on a dataset, as well against some of the currently available solutions.