PLAnCo: Provable Learnability Analysis of Generic APUF Compositions Using Finite Automata Models

Soumi Chatterjee, Durba Chatterjee, Aritra Hazra, Debdeep Mukhopadhyay

Abstract: Delay-based physically unclonable functions (PUFs) have been a popular candidate for hardware-based root-of-trust owing to their capability to implement “secret-free” cryptography in low-end devices. These are specialized circuits that leverage the intrinsic variabilities of their host device to realize a pseudorandom instance-specific Boolean function in hardware. Since the emergence of this primitive, there has been an ongoing make-and-break game where attackers propose novel attack strategies to model PUF constructions, and designers try to come up with fortified constructions to mitigate the state-of-the-art attacks. In this work, we present a formal learnability analysis framework based on automata-based model and prove that any APUF based PUF composition is PAC learnable. First, we introduce a formal framework for representing delay-based PUFs as deterministic finite automata (DFA), leading to a polynomial-sized representation of any arbitrary composition. Next, we establish the provable learnability of DFA compositions in the distribution-independent PAC model, achieving polynomial space and time complexity. To validate our theoretical findings, we conduct extensive experiments on simulated and FPGA implementations of PUF constructions using the PyPUF framework, thereby demonstrating the practicality/feasibility of our results.

External IDs:dblp:journals/tifs/ChatterjeeCHM25