Deep Learning Translation of Type-Unsafe Source and Binary Software

Programming languages have progressed significantly in recent years to support type/memory safety and concurrency at record execution speeds. Adoption of these new, safe languages has the potential to eliminate key exploitation categories such as heap and buffer overflows. The only holdup is the time, effort, and expertise needed to translate existing software into these preferable languages.   At the same time, deep learning is revolutionizing the machine language translation landscape. It has been embraced for production by giants like Google, and is being funded heavily for research in the domains of software decompilation and source code translation by companies like Facebook. This overlap in technological advancement provides a ripe opportunity to apply deep learning to programming language translation from unsafe legacy software to safe, new programming languages. We propose to explore this idea by implementing and evaluating the accuracy and performance of three such deep learning models. The model with the best results will be used to demonstrate the lifting of code in a medium-sized open source C/C++ code base project into the safe programming language Rust.