Co-authored by Yihua Liao and Yi Zhang

This is the fifth in a series of articles focused on AI/ML.  

Source code is a critical part of an organization’s intellectual property and digital assets. As more and more centralized source code repositories are moving to the cloud, it is imperative for organizations to use the right security tools to safeguard their source code.

In December 2020, a software engineer started working at Tesla and immediately began uploading the company’s source code files to his personal Dropbox account. Tesla didn’t confront him about his alleged theft until January 6, 2021. In March 2022, Microsoft confirmed that the Lapsus$ hacking group had compromised an employee account and stolen the company’s source code from Bing, Bing Maps and Cortana. These are just some of the latest examples of sensitive data leaking in the form of source code. 

Challenges of source code detection

It is not an easy task to determine programmatically whether a text document is source code or not. First of all, there are many different programming languages, and there is no specific pattern to describe what the source code should look like. As a result, it is impossible to come up with some regular expressions to match source code files with acceptable accuracy. 

Furthermore, programming languages are different from natural languages. Therefore, many popular pre-trained NLP (Natural Language Processing) models, such as GPT, BERT, and XLNet, which have shown great results in other document classification problems, are not effective in identifying source code. For example, some terms, punctuations, and symbols, such as “str”,  “def”, “==”, “>=”, and “:”, are not included in the vocabularies of most pre-trained models. However, they are widely used and carry significant meanings in source code. On the other hand, some words, such as “return” and “switch”, are used in both natural English and programming languages, yet with very different semantic meanings. 

The Netskope source code classifier

To address these challenges, we have developed a machine learning (ML) based source code classifier to detect source code files, as part of Netskope’s Advanced DLP (data loss prevention) solution. The source code classifier takes advantage of a proprietary code vocabulary, which consists of 80,000 common phrases in source code. The code vocabulary was extracted from a large corpus of source code sample files, covering more than 20 of the most popular programming languages. 

We have generated machine learning features based on the code vocabulary and trained a decision tree-like source code classifier. Compared to the model that is refined from a pre-trained language model, the source code classifier achieves 92% reduction in false positives while keeping the source code detection rate at 99%.   

The source code classifier scans our customers’ network traffic and looks for source code files inline. Its runtime in production is just a few milliseconds. This allows customers to enforce their source code policy and prevent data exfiltration in real time. 

More about Netskope DLP

Netskope’s award-winning DLP solution helps an organization protect the sensitive data it owns or its employees process. Netskope understands the context of cloud and web access, including the user, device, app, instance, activity, and content involved, to accurately identify violations and data risks. From there, it can then allow, challenge, block, quarantine, encrypt, or apply a legal hold, as well as integrate with on-premises solutions to prevent data loss and exposure. Netskope performs accurate inspection through 3,000+ out-of-the-box data identifiers, 25 predefined legal and regulatory compliance templates, and various matching techniques (proximity expression, custom regex and dictionaries, file fingerprinting, exact data matching, and so on).

Netskope Advanced DLP includes machine learning based file classification that provides a fast and effective way to identify sensitive documents, enabling users to work inline with granular real-time DLP policy controls. ML classifiers are able to accurately classify documents into different categories, including source code, tax forms, patent documents, and other sensitive legal and financial documents, without the need to identify specific pieces of sensitive information contained in those files. 

For more information, please check out our white paper Protecting Data Using Machine Learning.