Work

Exact Reference Bias in Byzantine-Robust Estimation: Finite-Sample Recovery, Contamination Floors, and Sensor-Record Stress Tests

Year: 2026
Type: preprint
Abstract: A trusted reference can anchor estimation when some workers are Byzantine, but a reference with known bias is not necessarily an unbiased guide to the truth. The decisive question is whether the direction of that bias can be recovered from the worker reports. We study scalar aggregation with an authenticated honest majority, independent centered bounded worker errors, and a finite trusted reference sample. We distinguish knowledge of the signed bias from knowledge of its magnitude alone. This distinction yields an exact-recovery threshold for a noiseless reference, a positive-risk impossibility result for noisy references, and a joint worker–reference information bound. We also prove that, when the unsigned bias is sufficiently small relative to a growing fixed Byzantine budget, the minimax risk approaches that of returning the exact reference itself. These results motivate Certified Exact-Reference-Bias Aggregation (CERBA), an a priori rule that selects among robust worker and reference-assisted estimates using explicit risk certificates. In 129,600 synthetic trials, CERBA’s worst-case certificate did not imply the lowest error under every simulated attack. A separate 1,500-trial analysis of recorded weather observations shows a similar dependence on the aggregation rule and attack, without claiming multi-device federated deployment. Together, the theory and experiments identify when exact bias information is useful, when it cannot overcome contamination, and why empirical accuracy must be distinguished from a uniform guarantee.
Language: English
Open Access status: green
OpenAlex ID: W7236570658
Cites:
Cited by:
Related to: 10