Shadmon, R. (CS) – Proximal Byzantine Agreement
Research on fault-tolerance protocols for approximate Byzantine agreement
(ABA) has largely focused on ensuring that distributed processes remain
consistent despite fewer than 1/3 faulty processes. Yet in many
real systems, consistency is only useful when it enables processes to
make accurate decisions from replicated, noisy, and potentially
adversarially corrupted data relative to an ideal fault-free baseline.
This limitation is increasingly important in edge applications such as
autonomous vehicles, drone networks, smart cities, manufacturing, and
sensor-based systems, where agreement directly drives downstream
actions. At the same time, many existing ABA protocols impose
impractical requirements, such as replica counts that grow with data
dimensionality or prior knowledge of the maximum distance between values
proposed by each process.
We introduce Stochastic Byzantine Agreement (SBA), a new problem
formulation in which the goal is to estimate an output from n replicated
values consisting of n-f nonfaulty outputs generated by an
underlying stochastic process and f arbitrarily chosen
Byzantine outputs. We then present Proximal Byzantine Agreement
(PBA), a stochastic agreement protocol that solves SBA by enabling
consumers to infer the most likely ideal output conditioned on the
outputs they receive. In addition, PBA provides a region
guarantee that, as we prove, always contains the corresponding
fault-free stochastic estimate of the true value.
We describe the design of PBA, formalize its guarantees, and evaluate
its accuracy against existing techniques using stochastic simulations
across symmetric and asymmetric distributions and multiple system
configurations. We also evaluate runtime overhead and performance in a
follow-the-leader drone network simulator and in a Java implementation on
Raspberry Pis using a real-world adaptive cruise control dataset. Our
results show that PBA performs competitively across all evaluated
settings and especially well under simulated Byzantine attack. Most
notably, PBA maintains stable accuracy as dimensionality increases,
outperforming methods that require up to 10x more replicas}
and incur up to 10x greater computation time per agreement
decision.
Event Host: Roy Shadmon, Ph.D. Candidate, Computer Science
Advisor: Owen Arden
Zoom: https://ucsc.zoom.us/j/98390167664?pwd=DwkNuUSRaZRKXYb7pQbDYXgf7HFFPg.1
Passcode: pba