Key points are not available for this paper at this time.
Physically unclonable functions (PUFs) have long been a key component of hardware-based device authentication. They rely on intrinsic manufacturing variability to give unique and tamper-resistant Identifiers for each silicon device. On the other hand, they have significant limitations, including high hardware overhead, aging degradation, and vulnerability to modeling and side-channel attacks. To address these restrictions, we previously presented virtual physically unclonable functions (VPUFs), a software-based solution that uses neural networks and split learning to improve scalability, flexibility, and deployment feasibility in resource-constrained Internet of Things (IoT) environments. Despite these advancements, VPUFs remain vulnerable to physical extraction and reverse engineering of the deployed model. In this paper, we present a lightweight, neuron-criticality-aware encryption framework that significantly enhances VPUF security. By conducting detailed ablation analysis, we identify the most critical neurons and selectively apply XOR-based encryption, minimizing computational overhead and preserving authentication accuracy. Coupled with a dynamic key generation mechanism based on Rayleigh fading through Jake's model, our approach achieves up to 99.4% added security with microsecond-scale latency.
Khan et al. (Fri,) studied this question.