Attack-Resistance and Reliability Analysis of Feed-Forward and Feed-Forward XOR PUFs

Abstract

Physical unclonable functions (PUFs) are lightweight hardware security primitives that are used to authenticate devices or generate cryptographic keys without using non-volatile memories. This is accomplished by harvesting the inherent randomness in manufacturing process variations (e.g. path delays) to generate random yet unique outputs. A multiplexer (MUX) based arbiter PUF comprises two parallel delay chains with MUXs as switching elements. An input to a PUF is called a challenge vector and comprises of the select bits of all the MUX elements in the circuit. The output-bits are referred to as responses. In other words, when queried with a challenge, the PUF generates a response based on the uncontrollable physical characteristics of the underlying PUF hardware. Thus, the overall path delays of these delay chains are random and unique functions of the challenge. The contributions in this thesis can be classified into four main ideas. First, a novel approach to estimate delay differences of each stage in MUX-based standard arbiter PUFs, feed-forward PUFs (FF PUFs) and modified feed-forward PUFs (MFF PUFs) is presented. Test data collected from PUFs fabricated using 32 nm process are used to learn models that characterize the PUFs. The delay differences of individual stages of arbiter PUFs correspond to the model parameters. This was accomplished by employing the least mean squares (LMS) adaptive algorithm. The models trained to learn the parameters of two standard arbiter PUF-chips were able to predict responses with 97.5% and 99.5% accuracy, respectively. Additionally, it was observed that perceptrons can be used to attain 100% (approx.) prediction accuracy. A comparison shows that the perceptron model parameters are scaled versions of the model derived by the LMS algorithm. Since the delay differences are challenge independent, these parameters can be stored on the server which enables the server to issue random challenges whose responses need not be stored. By extending this analysis to 96 standard arbiter PUFs, we confirm that the delay differences of each MUX stage of the PUFs follow a Gaussian probability distribution. Second, artificial neural network (ANN) models are trained to predict hard and soft-responses of the three configurations: standard arbiter PUFs, FF PUFs and MFF PUFs. These models were trained using silicon data extracted from 32-stage arbiter PUF circuits fabricated using IBM 32 nm HKMG process and achieve a response-prediction accuracy of 99.8% in case of standard arbiter PUFs, approximately 97% in case FF PUFs and approximately 99% in case of MFF PUFs. Also, a probability based thresholding scheme is used to define soft-responses and artificial neural networks were trained to predict these soft-responses. If the response of a given challenge has at least 90% consistency on repeated evaluation, it is considered stable. It is shown that the soft-response models can be used to filter out unstable challenges from a randomly chosen independent test-set. From the test measurements, it is observed that the probability of a stable challenge is typically in the range of 87% to 92%. However, if a challenge is chosen with the proposed soft-response model, then its portability of being stable is found to be 99% compared to the ground truth. Third, we provide the first systematic empirical analysis of the effect of FF PUF design choices on their reliability and attack resistance. FF PUFs consist of feed-forward loops that enable internally generated responses to be used as select-bits, making them slightly more secure than a standard arbiter PUFs. While FF PUFs have been analyzed earlier, no prior study has addressed the effect of loop positions on the security and reliability. After evaluating the performance of hundreds of PUF structures in various design configurations, it is observed that the locations of the arbiters and their outputs can have a substantial impact on the security and reliability of FF PUFs. Appropriately choosing the input and output locations of the FF loops, the amount of data required to attack can be increased by 7 times and can be further increased by 15 times if two intermediate arbiters are used. It is observed adding more loops makes PUFs more susceptible to noise; FF PUFs with 5 intermediate arbiters can have reliability values that are as low as 81%. It is further demonstrated that a soft-response thresholding strategy can significantly increase the reliability during authentication to more than 96%. It is known that XOR arbiter PUFs (XOR PUFs) were introduced as more secure alternatives to standard arbiter PUFs. XOR PUFs typically contain multiple standard arbiter PUFs as their components and the output of the component PUFs is XOR-ed to generate the final response. Finally, we propose the design of feed-forward XOR PUFs (FFXOR PUFs) where each component PUF is an FF PUF instead of a standard arbiter PUF. Attack-resistance analysis of FFXOR PUFs was carried out by employing artificial neural networks with 2-3 hidden layers and compared with XOR PUFs. It is shown that FFXOR PUFs cannot be accurately modeled if the number of component PUFs is more than 5. However, the increase in the attack resistance comes at the cost of degraded reliability. We also show that the soft-response thresholding strategy can increase the reliability of FFXOR PUFs by about 30%.