Privacy-friendly machine learning in practice

We believe that technology only becomes truly valuable when it contributes to people and society. That’s why we always look for ways to make existing systems smarter without collecting more data than necessary. Our project for 99gram.nl is a great example of that.

From forum data to valuable insights

The 99gram forum has been around for nearly 15 years. Beneath the surface lies a wealth of information: thousands of posts and replies carefully moderated over time. What makes this data special is that it already comes with human labels. A moderator has assessed whether each message was safe, inappropriate, or spam.

That makes this dataset ideal for use in a machine learning model. Not to collect more data, but to make better use of the reliable data that already exists.

A custom model instead of sharing data

Of course, we could outsource this task to a Large Language Model (LLM) like OpenAI. But that brings two main issues:

• Privacy: you might end up sharing sensitive data with an external party (outside the EU).
• High cost: there are recurring expenses involved.

That’s why we chose a different approach: training our own classification model, specifically for 99gram. This way, all data stays local, in a secure and controlled environment, fully in line with our ISO 27001 standards for information security.

Requirements for responsible training

Before we started, we defined strict requirements for both the dataset and the process:

• The data had to be clean: deleted or sensitive messages were filtered out.
• No personal information was allowed in the training data.
• The quality had to be high: 99gram’s moderation history made that possible.

This resulted in a naturally labeled training set with reliable categories:

• Safe: messages that are publicly visible online
• Toxic: messages that were removed by moderators
• Spam: messages that were blocked by automatic filters