Yimin Gao

Yimin Gao

CryptoPUF

IoT security has an awkward economics problem: the devices that most need protecting — battery-powered sensors, UHF RFID tags — are the ones with the least area and energy to spend on it. Physical unclonable functions are an appealing answer, since they derive a unique “digital fingerprint” from fabrication variation rather than from a key you have to store.

The problem is that machine learning got good. Modern ML models can learn the challenge-response behavior of conventional strong PUFs from collected CRPs, which defeats the point. Cryptographic algorithms don’t have this weakness — you can’t regress your way through a block cipher — but they have the opposite one: no intrinsic key, so you’re back to storing secrets on chip.

CryptoPUF puts a weak PUF together with a cryptographic encryption core so each covers for the other. Read one way, it’s a crypto-enhanced PUF: the crypto core shields the weak PUF from direct exposure while massively expanding the number of usable CRPs. Read the other way, it’s a PUF-enhanced crypto core with intrinsic key generation, which removes the need for on-chip key storage entirely.

The weak PUF never reaches the attackerA challenge enters a cryptographic core, which draws its key from a weak PUF held behind a boundary. Only the challenge and the crypto core's response are observable, so a machine-learning attacker never sees the PUF's own analog mapping. Weak PUF fabrication variation key Crypto core not ML-learnable on chip — never exposed response Attacker collects CRPs LR · SVM · MLP ~50% prediction accuracy no better than guessing Read the other way: a crypto core with an intrinsic key, so nothing has to be stored on chip.
The crypto core shields the weak PUF from direct exposure while multiplying the number of usable challenge–response pairs.

Against Logistic Regression, Support Vector Machine, and Multilayer Perceptron attacks it holds at near-ideal 50% prediction accuracy — attackers do no better than guessing — while keeping resource utilization low. Compared to published alternatives, it’s the most compact ML-resilient PUF we’re aware of.

Reliability

A PUF is only useful if it gives the same answer every time, across temperature and voltage corners and over the device’s life, and raw PUF bits don’t. My related circuit work designs delay-based CMOS PUFs at the transistor level with Monte Carlo variability analysis, and screens for the cells that stay stable — cheaply, and in a fully digital way.

Follow-on silicon is currently being taped out in GlobalFoundries 12nm.

Published at the IEEE International Conference on RFID Technology and Applications (RFID-TA), 2024. Work with J. Chilaka, E. Pantoja, R. Klenke, and M. R. Stan.