Augustine et al. prove robust distributed computing with majority adversarial workers
New paradigm achieves efficiency even when most workers are malicious.
In a new paper on arXiv, Augustine et al. tackle a fundamental challenge in distributed computing: how to reliably run parallel computations when a majority of workers might be malicious. They build on the recently introduced "supervised distributed computing" paradigm, where a reliable supervisor guides workers through an acyclic task graph. Unlike the classic master-worker model where the master verifies every result—creating a bottleneck—the supervisor here outsources verification to the workers themselves.
The key advance is that the authors prove robust and efficient solutions exist for any constant β < 1 (i.e., any fraction of adversarial workers, as long as it's strictly less than 100%). The expected work for honest workers is close to a single execution per task, thanks to a lightweight verification mechanism. This dramatically improves on prior robust master-worker and peer-to-peer methods, which required significantly more redundancy. The result opens the door to using untrusted volunteer computing pools (like BOINC) with strong guarantees, even if most participants are adversarial.
- Supervisor outsources task verification to workers, avoiding master bottleneck.
- Works for any constant fraction β<1 of adversarial workers (e.g., 51% malicious).
- Honest workers perform only ~1 execution per task expected—far less redundancy than prior approaches.
Why It Matters
Enables reliable, efficient parallel computing using untrusted volunteers or cloud nodes, even with majority adversaries.