Authors:

Kyung Min Ko

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Extracting necessary source code authorship attributes is crucial for successful identification. However, extracting such attributes presents significant challenges in real-world scenarios primarily due to various syntax rules in diverse programming languages, average code line availability, and a limited number of code samples per author. Initially, a common approach was to utilize source code to detect the author by extracting various features from source code such as design patterns and the name of the variables. Even though source code includes valuable features, often malware programs are only left with binary executables. Therefore, it is common to apply feature extraction for binary executables. Even though previous researchers developed solid solutions to solve the code authorship tasks, to the best of our knowledge, there are currently no work-related code language models. In our research, we are using the Code T5 model, which is capable of handling code- specific semantics. Common code language models have limitations on input token length, so instead of using the entire code, we leveraged functions present in the code. We used functions as input for the model, then combined the result to predict the author with majority votes. Furthermore, we applied contrastive learning, which learns useful representation from comparing similar and dissimilar dataset pairs to not only improve the accuracy but also deal with code with anonymous authors. We initially tested on 10 authors' datasets from Google Code Jam. Furthermore, we tested on the real-world malware dataset to expand our results. Our result demonstrates that predicting at the file level is also not robust and unstable, since we found the model mostly relies on functionality. Thus, we propose to predict at the function level and use majority voting.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Kyung Min Ko