Consensus when all processes may be Byzantine for some time

Abstract

a b s t r a c tAmong all classes of faults, Byzantine faults form the most general modeling of value faults. Traditionally, in the Byzantine fault model, faults are statically attributed to a set of up to t processes. This, however, implies that in this model a process at which a value fault occurs is forever ''stigmatized'' as being Byzantine, an assumption that might not be acceptable for long-lived systems, where processes need to be reintegrated after a fault.We thus consider a model where Byzantine processes can recover in a predefined recovery state, and show that consensus can be solved in such a model. Our model admits executions where over time every process is faulty as long as there are always enough correct processes.

Nonrational Belief Paradoxes as Byzantine Failures

“…When two processors disagree about the value of an input, this is merely the Lieutenants version of the problem (Lamport et al, 1982). Further, "processor" means nothing more than a peer agent in a parallel system (Lamport et al, 1982) or even a subsequent independent state of a single system (Biely & Hutle, 2009). Later papers on the Byzantine Generals problem thus often recast it in terms of "Byzantine Faults" which "present different symptoms to different observers" and "Byzantine Failures" in which systems requiring interactive consistency cannot achieve it due to Byzantine Faults (Driscoll, Hall, Paulitsch, Zumsteg, & Sivencrona, 2004).…”

“…Such processes will obviously be inconsistent with the correctly-functioning processes. Biely and Hutle (2009) call Byzantine Faults "arbitrary value faults" because the result is that there is no constraint on the output value of the process. Byzantine Faults are the most general model of faults because they do not assume that any degree of detection and correction is possible (Biely & Hutle, 2009).…”

Agreement Protocols on an Arbitrary Network in the Presence of a Mobile Adversary

“…In [2], a version of the problem is investigated under the additional assumption that no new processors can fail before recovered processors have successfully learned the current state of computation. With this modification, it is observed that successful protocols are possible for n > 3m.…”