Silicon physical unclonable functions (PUFs) are security primitives relying on intrinsic randomness of IC manufacturing. Strong PUFs have a very large input-output space which is essential for secure authentication. Several proposed strong PUFs use timing races to produce a rich set of responses. However, these PUFs are vulnerable to machine-learning attacks due to linear separability of the output function resulting from the additive nature of timing delay along timing paths. We introduce a novel strong silicon PUF based on the exponential current-voltage behavior in subthreshold region of FET operation which injects strong nonlinearity into the response of the PUF. The PUF, which we term subthreshold current array (SCA) PUF, is implemented as a two-dimensional nxk transistor array with all devices subject to stochastic variability operating in subthreshold region. Our PUF is fundamentally different from earlier attempts to inject nonlinearity via digital control techniques, which could also be used with SCA-PUF. Voltages produced by nominally identical arrays are compared to produce a random binary response.
SCA-PUF shows excellent security properties. The average inter-class Hamming distance, a measure of uniqueness, is 50.3%. The average intra-class Hamming distance, a measure of response stability, is 0.6%. Crucially, we demonstrate that the introduced PUF is much less vulnerable to modeling attacks. Using a machine-learning technique of support-vector machine with radial basis function kernel for best nonlinear learnability, we observe that \"information leakage\" (rate of error reduction with learning) is much lower than for delay-based PUFs. Over a wide range of the number of observed challenge-response pairs, the error rate is 3-35X higher than for earlier designs.