Privacy by Design: Move the Model, Not the Data
About this session
Most data leaks happen because we move data around. Every time data leaves its source to be trained on somewhere central, you create exposure you then have to spend time and money defending. There is another way. Keep the data where it is and send the model to it instead. That one decision changes how you build from the start, and it cuts exposure dramatically without giving up accuracy. In this talk I will show what that looked like in real work, where we reduced data exposure by 99.2%. I will walk through how we kept sensitive data local using federated learning, differential privacy, and layered aggregation, where the standard approaches fell apart, and the trade-offs we lived with. If you build AI that touches people's data, you will leave with a clear picture of how to keep that data where it belongs.
Speaker
Key takeaways
- Why data movement, not collection, is where exposure actually happens
- How to keep data local by moving the model to it, using federated learning
- Where layered privacy protection holds up in production, and where it breaks