Research On Strong PUF Resisting Machine Learning Attack Method

Strong physical Unclonable Function(PUF)is a promising lightweight hardware security primitive.When inputting a challenge,PUF utilizes process variation that is difficult to predict in the chip manufacturing process to output an unclonable response.Strong PUF can generate a large number of challenge response pairs(CRPs)which is very suitable for device authentication in a resource-constrained environment.However,attackers can model the PUF instance by collecting few CRPs.Therefore,PUF is vulnerable to machine learning attacks.To resist this attack,many defenses have been proposed.However,these defenses incur high hardware overhead,degenerate reliability and are inefficient against advanced machine learning attacks.In order to address these issues,we propose two anti-modeling attack structures——dynamic multi-key obfuscation structure and confgurable Tristate PUF structure——to resist machine learning attacks.This paper proposes a dynamic multi-key obfuscation structure,the core idea is that several stable responses are derived from the PUF itself and pre-stored as the obfuscation Keys in the enrollment phase,and then a true random number generator is used to select any two Keys to obfuscate challenges and responses respectively with the bitwise XOR operation.Besides,we also set up a dynamic update mechanism,when the number of CRPs collected by the attacker approaches the given threshold,the obfuscation Keys will be updated immediately.The obfuscation mechanism proposed in this paper obfuscates the mapping relationship between challenge and response with extremely low hardware overhead,and it can resist all current machine learning modeling attacks effectively.Experimental results show that for a 64×64 Arbiter PUF,when 32 obfuscation Keys are used,even if attackers collect 1 million CRPs,the prediction accuracy is still around 50% with Logistic Regression,Support Vector Machine,Artificial Neural Network,Convolutional Neural Network and Covariance Matrix Adaptive Evolution Strategy model,which is extremely low.Furthermore,through setting the threshold properly and updating the key dynamically,any machine learning attacks cannot break it successfully within the effective time.This paper proposes a configurable Tristate PUF structure,the core idea is to use the response generated by the arbiter PUF working state to XOR the challenge and response of the other two working states(ring oscillator PUF and bi-stable ring PUF).The reconfigurable Tristate PUF can resist machine learning attacks by hiding its working mode from attackers.We implemented a Tristate PUF structure on the Xilinx Artix-7 FPGA board.The experimental results show that the Tristate PUF structure can resist linear machine learning attacks effectively,such as logistic regression and nonlinear machine learning attacks,such as neural networks,while satisfying the requirements of uniformity,reliability and uniqueness.And the modeling accuracy of Tristate PUF is slightly higher than 50%,which is similar to random guessing.