Secure Aggregation Techniques
What you'll learn: How to combine model updates from multiple clients without exposing any individual client's data using cryptographic methods.
The Privacy Problem in Federated Learning
In federated learning (from lesson 1540), clients send model updates back to a central server for aggregation. But even without sharing raw data, these gradient updates can leak sensitive information about individual training examples. For instance, analyzing one client's weight changes might reveal whether a specific record was in their dataset.
Secure aggregation solves this: the server learns only the combined sum of all updates, never seeing any individual contribution.
How Secure Aggregation Works
Think of it like a secret ballot election where votes are tallied but individual votes remain hidden—except instead of votes, we're aggregating numerical gradient vectors.
The core cryptographic techniques include:
Additive Secret Sharing
Each client splits their update into random "shares" distributed to other clients. Only when all shares combine does the true aggregate emerge. Individual shares reveal nothing.
Threshold Encryption
Clients encrypt their updates such that decryption requires a minimum number of participants. The server can only decrypt the aggregate when enough clients contribute—never individual values.
Masking Protocol
Clients add cryptographic noise that cancels out in aggregate. Client A adds +random_mask, Client B adds -random_mask. The sum removes all masks while hiding individual contributions.
Why It Matters
Secure aggregation provides cryptographic privacy guarantees beyond differential privacy (lesson 1535). Even a malicious server administrator cannot inspect individual model updates—only the final aggregated result.
The trade-off? Increased computational overhead and communication rounds, plus vulnerability if too many clients drop out mid-protocol.
Key Takeaway: Secure aggregation uses cryptographic techniques like secret sharing and masking to aggregate federated model updates while provably preventing the server from seeing any individual client's contribution.