A Neural Decompiler for Java Classes

Anonymous

Abstract: We propose a novel approach using neural machine translation to automatically decompile entire Java classes. Our method relies only on {source code, bytecode} pairs of Java methods and does not require any additional domain knowledge of the target language. To overcome the token length limitations of current Transformer models, we partition class bytecode into methods, generate Java code for each method, and then reassemble all outputs into a final class. Our neural decompiler is able to generate more human-readable output (measured by CodeBLEU) than existing software-based decompilers while achieving slightly lower pass rates on fuzz tests. We will release our source code, dataset collection code, and pretrained Java class decompiler model to aid in development of more robust neural machine translators.

Paper Type: short

Research Area: Machine Translation