Secure Aggregation and Privacy Amplification
What you'll learn: How cryptographic protocols let federated learning systems combine model updates from many devices without exposing any individual participant's data.
The Privacy Problem in Federated Learning
In federated learning (covered in lesson 2843), devices train models locally and send updates to a central server. But even without sharing raw data, those gradient updates can leak private information through membership inference or model inversion attacks. The challenge: how do we aggregate everyone's contributions without the server—or anyone else—seeing individual updates?
How Secure Aggregation Works
Secure aggregation uses cryptographic techniques to ensure the server only learns the sum of all participants' model updates, never individual contributions.
Think of it like a salary survey: everyone wants to know the average salary, but nobody wants to reveal their own. Secure aggregation is like each person encrypting their salary in a special way so that when all encrypted values are combined, the encryption cancels out—revealing only the total sum, not individual amounts.
Common techniques include:
- Secret sharing: Each participant splits their update into random shares distributed among others
- Masking protocols: Participants add random masks that cancel out during aggregation
- Homomorphic encryption: Allows mathematical operations on encrypted data
Privacy Amplification Through Aggregation
Privacy amplification is the principle that combining many updates makes it exponentially harder to infer information about any single participant. The more participants in the aggregation, the stronger the privacy guarantee—even beyond what differential privacy alone provides.
When secure aggregation is combined with DP-SGD (lesson 2841), the privacy guarantees multiply: the server can't see individual updates, and even the aggregate reveals limited information.
Key Takeaway: Secure aggregation uses cryptographic protocols to compute aggregate model updates without revealing individual contributions, providing a crucial privacy layer that complements differential privacy in federated learning systems.