Sovereign Federated Learning with Byzantine-Resilient Aggregation

TMLR Paper6759 Authors

Abstract:

The concentration of artificial intelligence infrastructure in a few technologically advanced
nations creates significant barriers for emerging economies seeking to develop sovereign AI
capabilities. We present DSAIN (Distributed Sovereign AI Network), a novel federated
learning framework designed for decentralized AI infrastructure development in resourceconstrained
environments. Our framework introduces three key technical contributions: (1)
FedSov, a communication-efficient federated learning algorithm with provable convergence
guarantees under heterogeneous data distributions; (2) ByzFed, a Byzantine-resilient aggregation
mechanism that provides (ϵ, δ)-differential privacy while tolerating up to ⌊(n−1)/3⌋
malicious participants; and (3) a blockchain-based model provenance system enabling verifiable
and auditable federated learning. We provide theoretical analysis establishing convergence
rates of O(1/
√
T) for non-convex objectives and O(1/T ) for strongly convex objectives
under partial participation. Extensive experiments on CIFAR-10, CIFAR-100, and
real-world federated benchmarks demonstrate that DSAIN achieves accuracy within 2.3%
of centralized baselines while reducing communication costs by 78% and providing formal
privacy guarantees. We validate the framework through a deployment case study demonstrating
practical applicability in distributed computing environments.

Submission Type: Long submission (more than 12 pages of main content)

Changes Since Last Submission: First Submission

Assigned Action Editor: ~Eduard_Gorbunov1

Submission Number: 6759