Work

Prototype-Guided Defense in Edge AI: Securing Distributed Inference Against Model Poisoning in Resource-Constrained Systems

Year: 2026
Type: article
Abstract: The proliferation of edge artificial intelligence has enabled distributed inference across heterogeneous, resource-constrained devices, yet this architectural shift introduces acute vulnerabilities to model poisoning attacks that corrupt local updates or inference outputs. Traditional defense mechanisms, often designed for centralized cloud environments, impose prohibitive computational and communication overheads when adapted to the periphery of the network. This paper presents a systemic framework for prototype-guided defense, a technique that leverages representative class prototypes to detect and mitigate anomalous parameter deviations during distributed inference. We argue that prototype learning offers a structurally lightweight, intrinsically interpretable mechanism for securing edge AI systems without demanding substantial memory, bandwidth, or energy budgets. The exposition covers the threat model of targeted model poisoning in split inference and federated edge settings, the architectural incorporation of prototype consistency checks into local inference nodes, and the resulting trade-offs among security, accuracy, latency, and resource consumption. Broader implications for governance, sustainability, and fairness are examined, including the risk of prototype bias amplification and the need for standardized verification protocols. Through cross-domain analysis spanning autonomous vehicles, smart healthcare, and industrial IoT, we demonstrate that prototype-guided defenses, when carefully calibrated, can significantly strengthen the resilience of edge AI deployments. The paper concludes with forward-looking recommendations for policy frameworks and open research directions that balance security guarantees with the inherent constraints of embedded systems.
Language: English
Open Access status: diamond
OpenAlex ID: W7220046872
FWCI:
Cites: 11
Cited by:
Related to: 10