Course contentsShow
← Machine Learning and Deep Learning
Lesson 3358 of 3,53873. Privacy in Machine LearningFree lesson

Secure Aggregation Protocols

Using cryptographic techniques to aggregate updates without revealing individual client models.

Secure Aggregation Protocols

What you'll learn: How to combine model updates from multiple clients without any party—including the central server—seeing individual contributions.

The Core Problem

In federated learning, clients send gradient updates to a server for aggregation. But these updates can leak private information about training data. Even with differential privacy, the server still sees each client's update before adding noise or averaging. Can we aggregate updates without the server ever learning what any individual client contributed?

How Secure Aggregation Works

Secure aggregation uses cryptographic techniques to compute the sum (or average) of client updates while keeping individual contributions encrypted. Think of it like a secret ballot election where votes are counted without anyone seeing individual ballots.

The most common approach uses masking:

  1. Clients add cryptographic masks: Each client adds random noise to their update before sending it to the server
  2. Pairwise secret sharing: Clients agree on shared secrets with each other (not with the server) to generate these masks
  3. Masks cancel out: The masks are designed so that when all masked updates are summed, the random noise cancels perfectly, revealing only the aggregate

The server receives numbers like gradient + mask1 + mask2... from each client. When summed across all clients, the masks sum to zero by design, leaving only Σ gradients.

Why This Matters

Secure aggregation provides cryptographic privacy guarantees—the server literally cannot decrypt individual updates, even if it wanted to. Combined with differential privacy, this creates defense-in-depth: the server never sees raw updates, and the final aggregate has formal privacy guarantees.

Key Takeaway: Secure aggregation protocols use cryptographic masking to compute aggregate model updates without revealing any individual client's contribution, protecting privacy even from the central server.