AI-Native Secure and Resilient Edge Intelligence for Next-Generation IoT Networks: Vision, Metrics, and Research Roadmap

چکیده

Next-generation Internet of Things (IoT) networks are expected to support mission-critical, privacy-sensitive, and large-scale intelligent services under stringent constraints on latency, reliability, energy consumption, and trustworthiness. While edge intelligence has been widely advocated to overcome the limitations of cloud-centric processing, most existing designs remain “AI-assisted”, where learning and inference are treated as add-on functions and security is enforced through loosely coupled mechanisms. This separation becomes increasingly fragile in open and heterogeneous IoT ecosystems, where concept drift, adversarial behaviors, data poisoning, and device compromise can simultaneously undermine model performance and system security.This paper presents a vision for AI-native secure and resilient edge intelligence, in which intelligence is embedded into the network architecture and security is co-designed with learning and orchestration. We first identify the fundamental limitations of current IoT intelligence and security models, and then propose a set of performance metrics tailored to AI-native secure operation, including intelligence latency, adaptation speed under drift, privacy cost, attack robustness, and energy–security–accuracy tradeoffs. We further survey and systematize the enabling technologies required to realize this paradigm, spanning federated and split learning, privacy-preserving and trustworthy training, zero-trust collaboration, lightweight cryptography, and edge–cloud orchestration.Building on these foundations, we articulate key paradigm shifts from cloud-centric AI to edge-native intelligence, from accuracy-centric learning to resilience-aware learning, and from static security to adaptive and learning-integrated defense. Finally, we outline an open research roadmap covering architectural design, trustworthy model lifecycle management, cross-layer resource orchestration, and evaluation methodologies for realistic adversarial and non-stationary IoT environments. The proposed framework aims to guide the development of IoT systems where intelligence and security are not separate layers but jointly optimized properties of the network.