Fig. 5: 1-D HP strong memPUF resilient to machine learning attacks.
From: Halide perovskite memristors as flexible and reconfigurable physical unclonable functions
a Strong PUF architecture with a 1-bit response. Challenge consists of A bits for row selection (in this case, the first 3 bits are 1 and the last bit is 0) and log2(B/2) bits for column pair selection (here the first bit is 1). Natural current summation along columns gives rise to an exponential number of CRPs. b Machine learning (ML) results for strong PUF without recurrence. Accuracies close to 90% reveal the HP memPUF to be highly susceptible to such attacks. c Conceptual diagram of a recurrent PUF configuration shown as an unrolled cascade of PUFs without recurrence. After “n” steps of recurrence, the EN signal is turned on to stop further recurrence and enable the output bit R(n) to be readout. d ML results for strong PUF with recurrence. Accuracies reduce to almost wild guess probability with recurrence compared to c, proving that recurrence improves resistance to ML attacks.
