Inspiration
We were inspired by the scan-the-bank challenge. Adhering to GDPR is of utmost importance to safeguard our data privacy. Since this challenge is non-trivial and technically demanding, we decided to address it.
What it does
Our model can analyze a vast range of different file types (text, image, .pdf, etc.) and detect whether a file contains sensitive data.
How we built it
Our pipeline consists of a preprocessing, feature extraction, and a classification module. In the preprocessing module, files are specifically parsed based on their file type. For feature extraction, we employ a pretrained sentence transformer (all-MiniLM-L6-v2) to obtain semantically meaningful embeddings of each file. Finally, these embeddings are classified using a classical support vector machine.
Challenges we ran into
A major challenges was to make sure our final docker image does not grow bigger than 5 GB, which is particularly difficult when using libraries of large pretrained models, such as huggingface.
Accomplishments that we're proud of
That we managed to complete this task.
What we learned
We learned more about data privacy and we gathered hands-on experience with larger transformer models for NLP tasks
What's next for kevart-scan-the-bank
Implement more sophisticated preprocessing methods and leverage larger embedding models to further boost performance.
Built With
- docker
- huggingface
- python
- scikit-learn
- tesseract
Log in or sign up for Devpost to join the conversation.