Federated Uncertainty Consensus: Swarm-Level Safety Fallbacks for Disconnected Edge AI
Year: 2026
Type: preprint
Abstract: Edge Artificial Intelligence (AI) is increasingly deployed in latency-critical and highly dynamic environments, such as autonomous Unmanned Aerial Vehicle (UAV) swarms and disaster response robotics [3, 9]. A critical vulnerability in these systems is data drift—where the operational environment deviates from the training distribution, leading to silent model degradation [1, 2]. Current safety fallback mechanisms rely on hierarchical cloud architectures, where an edge device defers ambiguous inferences to a central server [4]. However, this is structurally impossible in disconnected, cloud-denied, or strictly localized ad-hoc networks. In this paper, we propose Federated Uncertainty Consensus (FUC), a novel technical AI safety protocol. Rather than defaulting to a centralized cloud, an edge node utilizes local Uncertainty Quantification (UQ) via Conformal Prediction [5]. Upon detecting out-of-distribution (OOD) anomalies, the node broadcasts a lightweight peer-to-peer query to the local swarm. The swarm mathematically aggregates its confidence, achieving a "safety fallback consensus" to override the degraded node. We outline the decentralized architecture, provide empirical latency metrics from a simulated ad-hoc swarm, and demonstrate that FUC maintains real-time safety thresholds without reliance on cloud infrastructure [7, 8].
Author: Christopher, Chidiebere
Institution: Caleb University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204872421
Cites:
Cited by:
Related to: