Course contentsShow
← Machine Learning and Deep Learning
Lesson 3372 of 3,53873. Privacy in Machine LearningFree lesson

Computational and Communication Costs

Understanding the overhead of cryptographic operations and communication rounds in secure aggregation systems.

Computational and Communication Costs

What you'll learn: You'll understand the practical overhead introduced by cryptographic operations and network communication in secure aggregation systems.

The Price of Privacy

Secure aggregation protects individual updates in federated learning through cryptographic techniques like secret sharing and masking. While these methods preserve privacy, they don't come free—there's a significant cost in both computation and communication.

Computational costs refer to the extra processing power needed for cryptographic operations. Each client must:

  • Generate random masks or secret shares
  • Encrypt their gradient updates
  • Perform modular arithmetic operations
  • Verify other participants' contributions (in some protocols)

Think of it like wrapping a gift: the actual gift (your model update) takes seconds to prepare, but the elaborate wrapping, ribbons, and tamper-proof seals might take 10x longer.

Communication costs measure the additional data transmitted over the network. Secure aggregation typically requires:

  • Multiple communication rounds (not just sending your update once)
  • Exchanging public keys or commitment values
  • Sending masked values that may be larger than raw updates
  • Broadcasting dropout/reconstruction information

Why This Matters

In federated learning with thousands of mobile devices, these costs multiply dramatically. A protocol requiring 3 communication rounds instead of 1 triples network usage. If cryptographic operations add 50% computational overhead per client, battery drain becomes a real concern on smartphones.

The key engineering challenge is balancing privacy guarantees against practical constraints: bandwidth limits, battery life, computation time, and monetary costs for cloud communication.

Key Takeaway: Secure aggregation protocols introduce measurable computational overhead from cryptographic operations and increased communication rounds—understanding these costs is essential for deploying privacy-preserving federated learning at scale.