- https://arxiv.org/pdf/2109.14236.pdf
- Source: arxiv
TL;DR
LightSecAgg replaces pairwise secret sharing in federated learning with a one-shot aggregate mask scheme, reducing secure aggregation overhead from O(n²) to O(n) while tolerating client dropout and Byzantine adversaries.
Summary
Standard secure aggregation in federated learning requires pairwise secret sharing between all clients, causing O(n²) communication and computation overhead that limits scalability. LightSecAgg replaces this with a one-shot encoding scheme where each client generates a single aggregate mask that cancels at the server, preserving privacy without revealing individual updates. The scheme scales linearly O(n), tolerates both client dropout and Byzantine participants within configurable thresholds, and integrates with standard FL protocols without requiring architectural changes.
Key contributions
- Proposes one-shot mask encoding as a replacement for pairwise secret sharing, achieving O(n) rather than O(n²) scaling.
- Demonstrates that aggregate masks that cancel at the server provide equivalent privacy to pairwise schemes.
- Shows that dropout tolerance can be built natively into the one-shot encoding without post-hoc recovery mechanisms.
- Establishes that Byzantine resistance is compatible with the lightweight encoding at configurable security thresholds.
When to cite
- When comparing secure aggregation protocols for federated learning and evaluating communication overhead.
- When arguing that O(n²) pairwise secret sharing is a scalability bottleneck in large FL deployments.
- When looking for a secure aggregation scheme that handles client dropout natively.
- When discussing linear-scaling cryptographic protocols as an alternative to quadratic multi-party computation designs.