Authors:

Shreya Ompreeti Ippili

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Code authorship attribution can be a groundbreaking step in Computer Science with numerous applications, ranging from solving plagiarism and authorship disputes to malware analysis and apprehension of those responsible. It also poses a threat to the privacy of programmers, necessitating the study of the "fingerprint" one leaves when they write code. Previous work has explored both source code and compiled binary attribution. While the results for source code are quite substantial, the attribution of compiled binary executables appears to be a lot more complex yet practical in the real world. Therefore, our study aims to leverage machine learning techniques for authorship attribution on compiled binary executables. Initially, we investigated whether pre-trained code language models such as CodeT5, when provided with source code from disassembled and decompiled binary executables, learn functionality over coding style. Training the model on datasets of different kinds with varying degrees of overlap in control flow and author contribution, we observed the test accuracies for each. Next, we plan on utilizing contrastive learning to attribute decompiled binary executables to authors by training the model to provide embeddings for the decompiled code, resulting in code from the same authors being closer in the embedding space. By comparing results from these different techniques, our work aims to determine the most effective approach for authorship attribution for decompiled binary executables and make progress based on the prior work that has been completed in this area.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Shreya Ompreeti Ippili