- https://inst.eecs.berkeley.edu/~cs261/fa18/scribe/10_15.pdf
- Source: berkeley
TL;DR
These lecture notes establish Shamir secret sharing as the foundational primitive for secure aggregation in federated learning under an honest-but-curious server threat model, covering the tradeoffs between communication overhead, dropout tolerance, and cryptographic security.
Summary
Federated learning requires aggregating model gradients from many clients without the server learning individual updates. These notes introduce Shamir secret sharing as the core mechanism: each client splits its gradient into shares distributed across peers, allowing the server to reconstruct only the aggregate gradient. The honest-but-curious server threat model defines the security boundary. Coverage includes the inherent tension between dropout tolerance (which requires a minimum share threshold) and security strength, plus communication overhead as a function of client count and sharing degree.
Key contributions
- Establishes the honest-but-curious server threat model as the standard baseline for secure aggregation analysis.
- Introduces Shamir secret sharing as the foundational primitive enabling private gradient aggregation in FL.
- Shows that dropout tolerance requires a minimum reconstruction threshold that trades off directly with security strength.
- Analyzes communication overhead as a function of client count and secret sharing degree, providing a scaling reference.
When to cite
- When introducing Shamir secret sharing as the basis for understanding secure aggregation protocols.
- When explaining the honest-but-curious threat model and its limitations for adversarial FL settings.
- When discussing the fundamental tradeoff between dropout tolerance and security guarantees in secret sharing.
- When establishing baseline communication cost for secure aggregation before comparing optimized protocols.