Sovereign-Mohawk: Machine-Checked Hierarchical Byzantine Resilience and Probabilistic Composition for Federated Learning

Abstract Federated learning (FL) at scale requires aggregating model updates across multi-tier topologies while tolerating adversarial participants. A common assumption in hierarchical FL is that if every local aggregation tier enforces an honest-majority guard, the composition inherits global Byzantine resilience. In this paper, we present Sovereign-Mohawk , an FL architecture and Lean 4 formal development that refutes and repairs this assumption. First, we present a machine-checked negative result: a concrete counterexample in Lean 4 showing that naive deterministic composition of per-tier Multi-Krum guards fails, allowing a 60% global Byzantine leaf fraction to pass local filters under concentrated placement. Second, we formalize a probabilistic repair proving that under explicit \(\:i.i.d.\) Bernoulli committee sampling, the probability of any tier exceeding its Byzantine filtering threshold is strictly bounded by a Chernoff bound over rationals, verified for \(\:T=200\) committees and \(\:c=50,000\) leaves per committee via native_decide (\(\:{P}_{\text{fail}}\le\:200\times\:{2}^{-1000}\)). Third, we couple these formal artifacts with a deployable Go runtime featuring Rényi differential privacy, flat Groth16/BN254 zero-knowledge verifiability, and post-quantum key exchange (X25519-ML-KEM-768). Finally, we evaluate the system across multi-node Kind (Kubernetes-in-Docker) clusters on single hosts (up to 500 pods), CI swarm matrices (500–1,500 nodes), and a 1,000-node simulator target under safe (10% Byzantine) and edge (35% Byzantine) adversarial regimes, demonstrating microsecond aggregation overhead (30.63 \(\:\mu\:\)s) and minimal accuracy impact (1.8% gap vs. clean baseline). Every formal claim is bound to implementation code via a publicly audited traceability matrix enforced in continuous integration. Federated Learning, Byzantine Fault Tolerance, Formal Verification, Lean 4, Hierarchical Aggregation, Differential Privacy, Zero-Knowledge Proofs.

Authors

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-11694611/v1
Primary Topic
Privacy-Preserving Technologies in Data
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Sovereign-Mohawk: Machine-Checked Hierarchical Byzantine Resilience and Probabilistic Composition for Federated Learning

Ryan Williams
Research Square
Privacy-Preserving Technologies in Data
preprint

Sovereign-Mohawk: Machine-Checked Hierarchical Byzantine Resilience and Probabilistic Composition for Federated Learning

Ryan Williams
preprint en

Abstract

Abstract Federated learning (FL) at scale requires aggregating model updates across multi-tier topologies while tolerating adversarial participants. A common assumption in hierarchical FL is that if every local aggregation tier enforces an honest-majority guard, the composition inherits global Byzantine resilience. In this paper, we present Sovereign-Mohawk , an FL architecture and Lean 4 formal development that refutes and repairs this assumption. First, we present a machine-checked negative result: a concrete counterexample in Lean 4 showing that naive deterministic composition of per-tier Multi-Krum guards fails, allowing a 60% global Byzantine leaf fraction to pass local filters under concentrated placement. Second, we formalize a probabilistic repair proving that under explicit \(\:i.i.d.\) Bernoulli committee sampling, the probability of any tier exceeding its Byzantine filtering threshold is strictly bounded by a Chernoff bound over rationals, verified for \(\:T=200\) committees and \(\:c=50,000\) leaves per committee via native_decide (\(\:{P}_{\text{fail}}\le\:200\times\:{2}^{-1000}\)). Third, we couple these formal artifacts with a deployable Go runtime featuring Rényi differential privacy, flat Groth16/BN254 zero-knowledge verifiability, and post-quantum key exchange (X25519-ML-KEM-768). Finally, we evaluate the system across multi-node Kind (Kubernetes-in-Docker) clusters on single hosts (up to 500 pods), CI swarm matrices (500–1,500 nodes), and a 1,000-node simulator target under safe (10% Byzantine) and edge (35% Byzantine) adversarial regimes, demonstrating microsecond aggregation overhead (30.63 \(\:\mu\:\)s) and minimal accuracy impact (1.8% gap vs. clean baseline). Every formal claim is bound to implementation code via a publicly audited traceability matrix enforced in continuous integration. Federated Learning, Byzantine Fault Tolerance, Formal Verification, Lean 4, Hierarchical Aggregation, Differential Privacy, Zero-Knowledge Proofs.

Research Square
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon