Secure VPUF Architectures for IoT Authentication: Split Learning, Watermarking, and Neuron-Aware Encryption

Abstract

The rapid expansion of IoT devices necessitates scalable, lightweight, and secure authentication mechanisms. This thesis proposes a robust software-based alternative to conventional hardware PUFs. The first contribution introduces a Virtual PUF (VPUF) framework that employs an encoder-decoder architecture within a split learning paradigm. By transmitting only latent representations between client and server, the proposed approach ensures efficient challenge-response generation while preserving model confidentiality. The second contribution enhances VPUF security through a deep learning-based digital watermarking scheme. A latent-space watermark, derived from Rayleigh fading characteristics via Jake’s model, is embedded into the VPUF response before transmission to add security during communication. This watermark provides both response concealment and a secondary layer of authentication, exhibiting high fidelity, strong noise resilience, and resistance to forgery. The third contribution presents a neuron-criticality-aware encryption mechanism. Through systematic ablation analysis, the most sensitive neurons in the encoder are identified and encrypted using lightweight XOR-based techniques to minimize physical access attacks. A dynamic key generation process, based on channel fading, further strengthens security by binding encryption keys to the wireless environment. Comprehensive evaluations of the watermarking and encryption modules demonstrate that the proposed VPUF maintains authentication performance, offers robust protection against tampering and parameter substitution, and incurs only microsecond-scale latency, making it well-suited for deployment in resource-constrained IoT settings.

Date of Award	2025
Original language	American English
Supervisor	Hani Saleh (Supervisor)