Improving Security and Reliability of Physical Unclonable Functions Using Machine Learning

Physical Unclonable Functions (PUFs) are promising security primitives for device authenti-cation and key generation. Due to the noise influence, reliability is an important performance metric of PUF-based authentication. In the literature, lots of efforts have been devoted to enhancing PUF reliability by using error correction methods such as error-correcting codes and fuzzy extractor. Ho-wever, one property that most of these prior works overlooked is the non-uniform distribution of PUF response across different bits. This wok proposes a two-step methodology to improve the reliability of PUF under noisy conditions. The first step involves acquiring the parameters of PUF models by using machine lear-ning algorithms. The second step then utilizes these obtained parameters to improve the reliability of PUFs by selectively choosing challenge-response pairs (CRPs) for authentication. Two distinct algorithms for improving the reliability of multiplexer (MUX) PUF, i.e., total delay difference thresholding and sensitive bits grouping, are presented. It is important to note that the methodology can be easily applied to other types of PUFs as well. Our experimental results show that the relia-bility of PUF-based authentication can be significantly improved by the proposed approaches. For example, in one experimental setting, the reliability of an MUX PUF is improved from 89.75% to 94.07% using total delay difference thresholding, while 89.30% of generated challenges are stored. As opposed to total delay difference thresholding, sensitive bits grouping possesses higher efficiency, as it can produce reliable CRPs directly. Our experimental results show that the reliability can be improved to 96.91% under the same setting, when we group 12 bits in the challenge vector of a 128-stage MUX PUF. Besides, because the actual noise varies greatly in different conditions, it is hard to predict the error of of each individual PUF response bit. This wok proposes a novel methodology to improve the efficiency of PUF response error correction based on error-rates. The proposed method first obtains the PUF model by using machine learning techniques, which is then used to predict the error-rates. Intuitively, we are inclined to tolerate errors in PUF response bits with relatively higher error-rates. Thus, we propose to treat different PUF response bits with different degrees of error tolerance, according to their estimated error-rates. Specifically, by assigning optimized weights, i.e., 0, 1, 2, 3, and infinity to PUF response bits, while a small portion of high error rates responses are truncated; the other responses are duplicated to a limited number of bits according to error-rates before error correction and a portion of low error-rates responses bypass the error correction as direct keys. The hardware cost for error correction can also be reduced by employing these methods. Response weighting is capable of reducing the false negative and false positive simultaneously. The entropy can also be controlled. Our experimental results show that the response weighting algorithm can reduce not only the false negative from 20.60% to 1.71%, but also the false positive rate from 1.26 × 10−21 to 5.38 × 10−22 for a PUF-based authentication with 127-bit response and 13-bit error correction. Besides, three case studies about the applications of the proposed algorithm are also discussed. Along with the rapid development of hardware security techniques, the revolutionary gro-wth of countermeasures or attacking methods developed by intelligent and adaptive adversaries have significantly complicated the ability to create secure hardware systems. Thus, there is a critical need to (re)evaluate existing or new hardware security techniques against these state-of-the-art attacking methods. With this in mind, this wok presents a novel framework for incorporating active learning techniques into hardware security field. We demonstrate that active learning can significantly im-prove the learning efficiency of PUF modeling attack, which samples the least confident and the most informative challenge-response pair (CRP) for training in each iteration. For example, our ex-perimental results show that in order to obtain a prediction error below 4%, 2790 CRPs are required in passive learning, while only 811 CRPs are required in active learning. The sampling strategies and detailed applications of PUF modeling attack under various environmental conditions are also discussed. When the environment is very noisy, active learning may sample a large number of mis-labeled CRPs and hence result in high prediction error. We present two methods to mitigate the contradiction between informative and noisy CRPs. At last, it is critical to design secure PUF, which can mitigate the countermeasures or modeling attacking from intelligent and adaptive adversaries. Previously, researchers devoted to hiding PUF information by pre- or post processing of PUF challenge/response. However, these methods are still subject to side-channel analysis based hybrid attacks. Methods for increasing the non-linearity of PUF structure, such as feedforward PUF, cascade PUF and subthreshold current PUF, have also been proposed. However, these methods significantly degrade the reliability. Based on the previous work, this work proposes a novel concept, noisy PUF, which achieves modeling attack resistance while maintaining a high degree of reliability for selected CRPs. A possible design of noisy PUF along with the corresponding experimental results is also presented