The Internet Integrity Initiative Team has introduced Piiranha-v1, a new model that helps detect and protect personal information online. As data privacy becomes more important, this model is a major step toward keeping sensitive information safe on various platforms and in multiple languages.

What is Piiranha-v1?

Piiranha-v1 is a small yet powerful model designed to find personally identifiable information (PII). Released under the MIT license, it supports six languages: English, Spanish, French, German, Italian, and Dutch. It can spot 17 different types of PII with an impressive 98.27% accuracy, making it a valuable tool for both businesses and individuals.

Top-Notch Detection

Built on the DeBERTa-v3 architecture, Piiranha-v1 excels at detecting various types of PII. It can find email addresses and passwords with 100% accuracy, helping to protect sensitive data. Even when it makes minor mistakes, like mixing up first and last names, it still correctly identifies the information as PII. This makes it highly useful in real-world situations where data isn't always perfectly organized.

How Was It Developed?

The team worked with partners like Hugging Face and Akash Network to develop Piiranha-v1. They trained the model on a huge dataset of over 400,000 records of masked PII, using H100 GPUs for speed and efficiency. The training process involved five rounds, using a batch size of 128. This careful training helped create a model that's both accurate and adaptable to different languages and contexts.

How Well Does It Work?

Piiranha-v1 shows strong results. When tested on around 73,000 sentences, it scored an F1 score of 93.12%. Its precision and recall rates are also high, at 93.16% and 93.08%, respectively. These numbers show that the model can accurately identify PII, even when the data is not in a standard format.

Where Can It Be Used?

The model is perfect for organizations that deal with a lot of personal data, such as banks, hospitals, and tech companies. By using Piiranha-v1, these organizations can automatically flag and hide sensitive information, helping to prevent data breaches and comply with privacy laws like GDPR and CCPA. The model is available on Hugging Face’s platform, making it easy to integrate into existing systems.

A Word of Caution

While Piiranha-v1 is very accurate, the developers advise using it carefully. Like all machine learning models, it's not perfect and may make mistakes, especially in such a complex task as PII detection across various languages. It's a strong tool but should be part of a broader strategy for data privacy.

How to Get Piiranha-v1?

Piiranha-v1 is available under the MIT license, allowing for wide use, including commercial purposes. By making it open-access, the Internet Integrity Initiative Team aims to improve data privacy worldwide. This means more organizations can protect personal information and reduce the risk of data breaches.

Conclusion

Piiranha-v1 is a big step forward in finding and protecting personal information online. Its high accuracy, ability to work in multiple languages, and flexibility make it a must-have tool for any organization looking to boost its data privacy efforts. As concerns about digital privacy grow, tools like Piiranha-v1 will play a key role in keeping sensitive information safe.