TABLNet: Trustworthy Adaptive Broad Learning Network for Physical-Layer Device Authenticationin Intelligent IoT Networks

Abstract Securing the expanding internet of things (IoT) ecosystem demands lightweightauthentication mechanisms compatible with hardware-constrained edge end-points. Specific emitter identification (SEI) exploits manufacturing-induced radiofrequency (RF) fingerprints to verify device identity at the physical layer, cir-cumventing the overhead of upper-layer cryptographic schemes. Nevertheless,contemporary deep generative architectures employed for automated RF fin-gerprint synthesis impose prohibitive computational demands that hinder theirdeployment on resource-limited platforms. In this paper, a trustworthy adap-tive broad learning network (TABLNet) framework is proposed to overcomethese limitations. Unlike deep hierarchical models, TABLNet adopts a lat-erally expandable flat architecture wherein raw signal samples are projectedonto generative feature nodes directly connected to the decision layer. Weightoptimization relies on pseudoinverse computation rather than iterative back-propagation, thereby suppressing gradient instability and substantially curtailingtraining duration. Moreover, an autonomous node expansion mechanism dynam-ically calibrates the network width, eliminating manual hyperparameter search.Comprehensive experiments on two hardware-captured RF datasets demonstratethat TABLNet achieves competitive or superior classification accuracy whiledelivering an order-of-magnitude reduction in training latency compared withprevailing deep learning baselines.