In the continuous battle to enhance threat detection and prediction measures, South Korean researchers have made a significant leap forward with the development of DarkBERT, a cutting-edge language model explicitly designed to decipher the murky language environment of the Dark Web. This innovation carries immense potential for revolutionizing threat intelligence gathering, giving organizations an advantage to stay a step ahead of evolving cyber threats.


The Dark Web is a part of the internet not indexed by search engines. Due to its anonymous nature, it requires access to special software (like Tor), often associated with illegal activities. Conversely, the Surface Web is the part of the web that is open to the public and indexed by conventional search engines such as Google. Lastly, BERTCoDA and BERTReddit are the BERT language model using CoDA and Reddit data sets.

Ethical Considerations

While the Dark Web contains significant unethical content, South Korean developers have committed to strong ethical guidelines. Ensuring sensitive or potentially illicit content, often associated with the Dark Web, was meticulously removed during the data processing stage. This moral commitment ensured a responsible approach to utilizing Dark Web data. Furthermore, to protect individual privacy, any sensitive information, such as personal emails, phone numbers, or IP addresses, that might have inadvertently been captured during the data-gathering process was conscientiously masked. The team also took proactive steps to store only raw text data, carefully filtering out non-text media.

DarkBERT Development

The development of DarkBERT, a tool used to understand and generate human language (a field known as Natural Language Processing or NLP), unfolded in four steps. First, it began with the same initial building blocks as a larger model, BERT. Then, DarkBERT was taught to reproduce the sophisticated behavior of BERT while being a more straightforward and less demanding model. This step is known as distillation. The third step was akin to carefully trimming a tree, pruning away some connections, and adjusting others within DarkBERT, which helped to make the model more efficient while preserving its effectiveness. Lastly, it was fine-tuned to ensure DarkBERT could perform well in real-world tasks, like understanding the context of the text or identifying specific elements in a sentence. This process involved giving DarkBERT additional practice on tasks it would encounter in its role.

Despite being less complex, DarkBERT continued to perform at a high level, thanks to its ability to learn further from specific types of data it would be working with. One area where DarkBERT excels is its proficiency in detecting and identifying keywords tied to illegal activities. Picture DarkBERT as a 'virtual bloodhound' capable of sniffing out potential threats in an extensive maze of text data. The effectiveness of DarkBERT in this task was measured using a metric known as Precision at k (P@k), essentially a gauge of how accurately DarkBERT can highlight the 'bad apples' among its top 'k' predictions.

To better grasp the significance of DarkBERT's performance, consider the comparison to two other models: BERTCoDA and BERTReddit. First, examine each model's 'bad apple' detection rate in their top 10, 20, 30, 40, and 50 predictions. At the top 10 and 20 levels, DarkBERTCoDA scored a 0.60 P@k, outpacing BERTCoDA and BERTReddit, which scored 0.40 each. At the top 30 levels, DarkBERTCoDA scored 0.50, slightly behind BERTReddit at 0.60 but on par with BERTCoDA. However, when we reach the top 40 and 50 predictions, DarkBERTCoDA maintains a consistent score of 0.42, demonstrating its steady performance.

These findings illustrate that DarkBERT is not only capable but, in many cases, superior to other models when it comes to identifying between 10 and 20 related keywords. This positions DarkBERT as a powerful tool in threat detection, providing organizations the means to proactively identify and neutralize potential threats lurking within the Dark Web.


DarkBERT demonstrated an impressive knack for identifying Dark Web-specific dialects, even though it occasionally missed synonyms frequently used on the Surface Web. A stark example is seen in how DarkBERT deals with terms like "Tesla" and "Champagne.” To a layperson or even standard language models, these words might denote a car brand or a type of wine respectively. However, on the Dark Web, they are often used as codewords for certain drugs.

Despite these advancements, DarkBERT has its share of challenges. It depends heavily on the availability of task-specific data from the Dark Web, which isn't freely available. This means the researchers might need to manually annotate or generate necessary data to take full advantage of DarkBERT's capabilities. Plus, DarkBERT's design is primarily optimized for English text, reflecting the fact that most Dark Web content is in English. For non-English tasks, additional training with language-specific data may be needed.


DarkBERT represents a huge stride towards leveraging the Dark Web for good, particularly in gaining better threat intelligence. The developers plan future enhancements, like integrating newer architectures and accumulating more data for a multilingual model. But even in its current form, DarkBERT is a robust tool with immense promise for future threat detection on the Dark Web.


DarkBERT: A Language Model for the Dark Side of the Internet

By Joshua Ivy, Information Security Analyst

Joshua is a new addition to the TraceSecurity team, bringing with him a wealth of experience from 20 years of service in the US Navy, with his last two years spent as an ISSM in Virginia Beach. He currently holds multiple industry certifications, most notably, CompTIA Security+, Pentest+, CySA, and is looking forward to graduating with a Bachelor's in Cybersecurity Technologies by the end of 2024. At TraceSecurity, he primarily focuses on penetration tests, risk assessments, and IT security audits.