Course contentsShow
← Security
Lesson 2843 of 3,05253. AI and ML SecurityFree lesson

Federated Learning Privacy

Privacy considerations when training models across distributed data sources without centralizing sensitive information.

Federated Learning Privacy

What You'll Learn

You'll learn how federated learning protects sensitive data by training models on distributed devices while addressing unique privacy risks that still emerge from sharing model updates.


What Is Federated Learning Privacy?

Federated learning solves a critical privacy challenge: training powerful ML models without collecting raw data into a central location. Instead of sending your photos, messages, or health data to a server, your device trains a local model copy and sends only model updates (gradients or weights) back to a central coordinator.

Think of it like collaborating on a recipe. Instead of sending everyone's secret ingredients to a central kitchen, each chef improves the recipe at home and shares only "add more salt" or "reduce sugar" — the tweaks, not the ingredients themselves.

The Privacy Challenge

While federated learning avoids centralizing raw data, model updates themselves leak information:

  • Gradient analysis can reveal training samples (similar to model inversion attacks you've studied)
  • Repeated participation enables membership inference — attackers detect if specific data was in your local training set
  • Aggregation points become attack surfaces where adversaries might observe individual updates before they're combined

Privacy Protection Strategies

Building on differential privacy concepts you already know:

  1. Local differential privacy: Each device adds calibrated noise to its gradients before sharing (like DP-SGD applied locally)
  2. Secure aggregation: Cryptographic protocols ensure the server only sees the sum of updates, never individual contributions
  3. Update clipping and sampling: Limit gradient magnitudes and randomly select participants per round
  4. Fewer communication rounds: Reduce how often devices share updates to minimize exposure

The privacy-utility tradeoff appears again: stronger privacy (more noise, fewer rounds) typically means slower convergence or reduced accuracy.

Key Takeaway:
Federated learning keeps raw data decentralized but introduces new privacy risks through model update leakage — combining differential privacy, secure aggregation, and careful protocol design is essential to protect participants' sensitive information.