Machine learning based strong puf stable challenge selection method

Abstract

The application discloses a strong PUF stable excitation screening method based on machine learning, which quantifies the stability of a group of excitations by using 1 and 0, defines the group of excitations with a stability of 1 as stable excitation and the group of excitations with a stability of 0 as unstable excitation, defines the stability of a group of excitations and its corresponding response as CSP, then uses the CSP to construct a machine learning training data set to train a machine learning model to obtain a screening model, thereby associating the excitation stability with the screening model, and in the actual application of the strong PUF, the excitation stability of the strong PUF can be determined through the screening model, so that the stable excitation of the strong PUF is screened to form a stable excitation set as the input of the strong PUF for identity information extraction; the advantage is that the stability of various strong PUFs can be effectively improved, and the stronger the external interference is, the more obvious the effect is.

Description

TECHNICAL FIELD

[0001] The application relates to a strong PUF stable challenge selection method, in particular to a strong PUF stable challenge selection method based on machine learning. BACKGROUND

[0002] With the development of the Internet of Things (IoT), all things perception and all things interconnection have become the main trend of development in various fields of production and life. However, along with it comes various hardware and software attack techniques against Internet of Things devices. Physical unclonable function (PUF) uses the physical differences of the circuit itself to extract a unique identifier, which can be widely applied in device authentication and key generation and other fields. Its lightweight, non-storage and non-volatile functions have broad application prospects in solving various Internet of Things security problems. The input of PUF is called challenge, and the output is called response. PUF generates a corresponding response output for each input challenge, and each challenge and its corresponding response form a challenge response pair (CRP). According to the different CRP generation capabilities, PUF is divided into strong PUF and weak PUF. Since the number of CRP of strong PUF increases exponentially with the expansion of the structure, it is more suitable for device authentication. It is also because of the huge number of CRP of strong PUF that it is impossible to test the stability of the corresponding response of all challenges in a short time. Its unstable bits will bring serious hidden dangers in high-precision and time-sensitive applications such as smart medical treatment and unmanned driving.

[0003] The device authentication protocol based on strong PUF mainly includes a registration phase and an authentication phase, and the device ID information can be directly provided by the CRP or provided by the PUF model. Due to the influence of environmental noise, transmission error and other factors in the authentication process, it is difficult to achieve 100% matching of the information. Therefore, a threshold value epsilon is set in the protocol, and when the matching degree exceeds the threshold value epsilon, the authentication is considered successful. In the authentication process, if a group of challenges are randomly selected for verification, the response generated by the unstable challenge will cause the authentication matching degree to decrease, and only the threshold value epsilon can be reduced to ensure the verification to pass, resulting in a doubling of the security risk of the protocol. In terms of machine learning modeling attacks against arbiter PUF (APUF), the number of CRP required to reduce the prediction accuracy from 98% to 95% is reduced by more than 10 times, that is, the attacker only needs to collect one-tenth of the CRP to simulate the interaction between the device and the server. If the server is subjected to a brute force attack, if 100 groups of challenges are selected for verification, the number of attacks required to reduce the threshold value from 98% to 95% is reduced from 2.51x10 26 to 1.597x10 22, which is 15000 times less, which seriously threatens the security of strong PUF authentication protocols.

[0004] Currently, in order to enhance the stability of PUF and exclude the influence of unstable bits on the overall structure, researchers have proposed a variety of solutions. Commonly used techniques include error-correcting code (ECC), temporal majority voting (TMV), and automatic self-checking and healing (ASCH). However, these techniques are all proposed for weak PUF, which generates information from a PUF module, and strong PUF, which generates information from multiple PUF modules. In weak PUF, unstable units can be screened or deleted, and in strong PUF, stability is determined by multiple modules, so it is not possible to improve stability by changing a specific module. These techniques cannot be applied to strong PUF. In addition, some researchers have proposed increasing error correction circuits to enhance the stability of strong PUF. Although this scheme can indeed enhance the stability of strong PUF, it often involves a large area overhead and is not suitable for IoT node applications.

[0005] Therefore, it is of great significance to design a strong PUF stability incentive screening method based on machine learning that can enhance the stability of strong PUF without increasing the area overhead. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a strong PUF stability incentive screening method based on machine learning that can significantly improve the stability of strong PUF, increase the stability of strong PUF to nearly 100% ideal value, and has less influence on strong PUF under external interference, without increasing the area overhead.

[0007] The technical solution adopted by the present application to solve the above technical problem is: a strong PUF stability incentive screening method based on machine learning, comprising the following steps:

[0008] Step 1, randomly generate n groups of incentives to test the stability of strong PUF, n is an integer greater than or equal to 100000, input each group of incentives in n groups of incentives into strong PUF m times, m is an integer greater than or equal to 100, obtain m responses corresponding to each group of incentives, if the m responses corresponding to a group of incentives in n groups of incentives are all the same, then the group of incentives is a stable incentive, and the stability of the corresponding response is 1, otherwise the group of incentives is an unstable incentive, and the stability of the corresponding response is 0, thereby obtaining the stability of the response corresponding to each group of incentives in n groups of incentives;

[0009] Step 2, define the stability of each group of challenges and its corresponding response as a CSP (Challenge-Stability Pair), and each group of challenges in n groups of challenges and its corresponding response constitutes a CSP, and n groups of CSPs are obtained, and the n groups of CSPs are used to form a machine learning training data set;

[0010] Step 3, obtain a machine learning model, adjust the number of parameters to be solved in the machine learning model to be the same as the number of bits of a group of challenges of the strong PUF, and each parameter to be solved corresponds to a bit of a group of challenges; each group of challenges in each CSP is used as the input of the machine learning model, and the stability of the corresponding response of the group of challenges is used as the output of the machine learning model, and the n groups of CSPs are used to train the machine learning model until the machine learning model converges, at which time the machine learning model training is completed, and the trained machine learning model is the screening model, wherein, in the process of training the machine learning model, an optimization algorithm is used to optimize the minimum value of the loss function of the machine learning model, and when the loss function is minimized, it indicates that the machine learning model converges;

[0011] Step 4, when the strong PUF is actually applied, the screening model is used to screen stable challenges, and the specific process is as follows:

[0012] S1, randomly generate multiple groups of challenges of the strong PUF, and use the multiple groups of challenges to form an initial challenge set, and the multiple groups of challenges in the initial challenge set are randomly sorted;

[0013] S2, sequentially input the multiple groups of challenges in the initial challenge set into the screening model as the input of the screening model, and the output of the screening model is the stability of the group of challenges corresponding to each input group of challenges, if the output of the screening model is 1, it indicates that the group of challenges input into the screening model is a stable challenge; if the output of the screening model is 0, it indicates that the group of challenges input into the screening model is an unstable challenge; before all the multiple groups of challenges in the initial challenge set are input into the screening model or after all the multiple groups of challenges in the initial challenge set are input into the screening model, if the number of stable challenges obtained meets the actual application requirement, stop inputting; after all the multiple groups of challenges in the initial challenge set are input into the screening model, if the number of stable challenges obtained still does not meet the actual application requirement, return to step S1 until the number of stable challenges obtained meets the actual application requirement;

[0014] Step 5, use all the stable challenges determined in step 4 to form a stable challenge set, and the stable challenge set is used as the input of the strong PUF for identity information extraction.

[0015] The machine learning model is any one of a support vector machine (SVM) model, an artificial neural network (ANN) model, a convolutional neural network (CNN) model and a light gradient boosting machine (LightGBM) model.

[0016] The optimization algorithm is any one of a gradient descent method, a Newton method and a swarm intelligence algorithm.

[0017] Compared with the prior art, the present application has the advantages that by adopting 1 and 0 to quantify the stability of a set of stimuli, defining the set of stimuli with a stability of 1 as stable stimuli and the set of stimuli with a stability of 0 as unstable stimuli, defining the stability of a set of stimuli and its corresponding response as a CSP (Challenge-Stability Pair), then adopting the CSP to construct a machine learning training data set to train a machine learning model to obtain a screening model, the stimulus stability is associated with the screening model, and in the actual application of the strong PUF, the stimulus stability of the strong PUF can be determined through the screening model, so that stable stimuli of the strong PUF are screened out to form a stable stimulus set as the input of the strong PUF for identity information extraction, the present application models the relationship between the stability of a stimulus and its corresponding response by a machine learning method, and the stability of any stimulus is calculated through the screening model, so that the stimuli in the random stimulus set can be screened before the random stimulus set is input to the strong PUF, the unstable stimuli are discarded, and the stable stimuli are screened out, so that the stable stimuli are determined before the strong PUF is tested, thereby the stability of the strong PUF can be greatly improved without additional error correction circuits, the stability of the strong PUF is improved to close to 100% ideal value, there is no risk of information leakage, and since there is a function corresponding relationship between the input and output of the strong PUF, the present application is suitable for any strong PUF and has strong universality, so that the present application can greatly improve the stability of the strong PUF, the stability of the strong PUF is improved to close to 100% ideal value, the strong PUF is less affected by external interference, and the area overhead is not increased, and experimental results show that the method of the present application can effectively improve the stability of various strong PUFs based on APUF, and the stronger the external interference, the more obvious the effect. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 It is a schematic diagram of several existing strong PUF structures based on APUF;

[0019] Figure 2 It is a flow principle diagram of the strong PUF stable stimulus screening method based on machine learning of the present application;

[0020] Figure 3 FIG. 4(a) is a response distribution diagram corresponding to an initial stimulus set screened by the strong PUF stable stimulus screening method based on machine learning of the present application;

[0021] FIG. 4(a) is a response distribution diagram corresponding to an initial stimulus set screened by the strong PUF stable stimulus screening method based on machine learning of the present application;

[0022] FIG. 4(b) is a response distribution diagram corresponding to a stable stimulus set obtained after screening the initial stimulus set by the strong PUF stable stimulus screening method based on machine learning of the present application;

[0023] FIG. 5(a) is an autocorrelation coefficient distribution diagram of 100000 responses corresponding to an initial stimulus set screened by the strong PUF stable stimulus screening method based on machine learning of the present application;

[0024] FIG. 5(b) is an autocorrelation coefficient distribution diagram of 100000 responses corresponding to a stable stimulus set obtained after screening the initial stimulus set by the strong PUF stable stimulus screening method based on machine learning of the present application; DETAILED DESCRIPTION

[0025] The present application will be further described in detail below with reference to the accompanying drawings.

[0026] Embodiment: As shown in the figure, a strong PUF stable stimulus screening method based on machine learning includes the following steps: Figure 2

[0027] Step 1, randomly generate n groups of stimuli for testing the stability of the strong PUF, n equals 1000000, input each group of stimuli in the n groups of stimuli into the strong PUF m times, m equals 100, obtain m responses corresponding to each group of stimuli, if the m responses corresponding to a group of stimuli in the n groups of stimuli are all the same, then the group of stimuli is a stable stimulus, and the stability of the corresponding response is 1, otherwise the group of stimuli is an unstable stimulus, and the stability of the corresponding response is 0, thereby obtaining the stability of the response corresponding to each group of stimuli in the n groups of stimuli;

[0028] Step 2, define a group of stimuli and the stability of the corresponding response as CSP (Challenge-Stability Pair), and form a CSP with each group of stimuli and the stability of the corresponding response in the n groups of stimuli, thereby obtaining n groups of CSPs, and use the n groups of CSPs to form a machine learning training data set;

[0029] ​Step 3, obtain a machine learning model, adjust the number of parameters to be solved in the machine learning model to be the same as the number of bits of a set of stimuli of the strong PUF, each parameter to be solved corresponds to a bit of stimulation in a set of stimuli; a set of stimuli in each CSP is used as the input of the machine learning model, and the stability of the corresponding response is used as the output of the machine learning model, and n sets of CSPs are used to train the machine learning model until the machine learning model converges, at which time the machine learning model training is completed, and the trained machine learning model is the screening model, wherein, in the process of training the machine learning model, an optimization algorithm is used to optimize the minimum value of the loss function of the machine learning model, and when the loss function is minimized, it indicates that the machine learning model converges;

[0030] Step 4, when the strong PUF is actually applied, the screening model is used to screen stable stimuli, and the specific process is:

[0031] S1, randomly generate multiple sets of stimuli of the strong PUF, and use the multiple sets of stimuli to form an initial stimulus set, and the multiple sets of stimuli in the initial stimulus set are randomly sorted;

[0032] S2, sequentially input the multiple sets of stimuli in the initial stimulus set into the screening model as the input of the screening model, and the output of the screening model is the stability corresponding to the set of stimuli input into the screening model. If the output of the screening model is 1, it indicates that the set of stimuli input into the screening model is a stable stimulus. If the output of the screening model is 0, it indicates that the set of stimuli input into the screening model is an unstable stimulus. Before all the multiple sets of stimuli in the initial stimulus set are input into the screening model or after all the multiple sets of stimuli in the initial stimulus set are input into the screening model, if the number of stable stimuli obtained meets the actual application requirement, the input is stopped. After all the multiple sets of stimuli in the initial stimulus set are input into the screening model, if the number of stable stimuli obtained still does not meet the actual application requirement, return to step S1 until the number of stable stimuli obtained meets the actual application requirement.

[0033] Step 5, all the stable stimuli determined in step 4 form a stable stimulus set, which is used as the input of the strong PUF for identity information extraction.

[0034] In this embodiment, the machine learning model is any one of a support vector machine (SVM) model, an artificial neural network (ANN) model, a convolutional neural network (CNN) model, and a light gradient boosting machine (LightGBM) model.

[0035] In this embodiment, the optimization algorithm is any one of a gradient descent method, a Newton method, and a swarm intelligence algorithm.

[0036] To verify the feasibility of machine learning for modeling the screening model of CSP correspondence, 25000 groups of CSP of 6XOR-APUF were collected to respectively test the modeling accuracy of the LightGBM model and the SVM model under different numbers of training data sets, and the experimental results are as shown in Figure 3 Figure 3 It can be seen that the modeling accuracy of the LightGBM model and the SVM model is close to the ideal value of 100% when the number of CSP in the training data set reaches 20000, proving that machine learning is suitable for modeling the CSP correspondence. And the LightGBM model has higher modeling accuracy and requires fewer CSPs than the SVM model in the method of the present application, and is the preferred machine learning model of the present application.

[0037] The definition of stability is the proportion of stable bits to the total response. As a statistical quantity, its accuracy is greatly affected by the change of sample size. The larger the amount of excitation data used to test stability, the more accurate the stability value obtained. When the statistical quantity is greater than 4000, the stability changes little, so the statistical quantity should be greater than 4000 when the stability of the strong PUF is counted. The noise coefficient represents the strength of the external interference on the strong PUF. The larger the noise coefficient, the more unstable the signal in the transmission process, and the worse the stability of the strong PUF. In the screening model construction stage of the present embodiment, one million groups of CSP data are collected to train the machine learning model, and the machine learning model is a LightGBM model with an optimization iteration number of 100,000. The initial excitation set data used for screening is 100,000 to ensure that there are enough excitation for testing the stability of the strong PUF after screening. The schematic diagrams of several existing APUF-based strong PUF structures are as shown in Figure 1 Figure 1 The APUF, 3XOR-APUF and 4-MPUF shown in

[0038] Table 1

[0039]

[0040] In Table 1, random represents the initial excitation set corresponding data before screening, and after screening, it represents the stable excitation set corresponding data obtained by screening the initial random excitation set using the method of the present application.

[0041] ​​From the analysis of Table 1, when the noise coefficients of the three strong PUFs are 0.2, 0.125 and 0.1 respectively, after screening, the stability of APUF is improved from 57.143% to 95.752%, 3XOR-APUF is improved from 46.264% to 78.263%, and 4-MPUF is improved from 61.708% to 67.528%. The stability of the above three types of strong PUFs after screening by the method of the application is significantly improved, and the lower the stability, the better the improvement effect. Since the accuracy and the complexity of the function of the machine learning modeling between the stimulus and the stability are inversely proportional, the stability after screening is related to the complexity of the strong PUF. Through the comparison of the prediction rate after screening of different strong PUFs, it can be found that the simpler the PUF structure, the higher the stability after screening. APUF is the simplest among the other two PUF structures, and the final stability is also the highest. In addition, 32-bit strong PUF is easier to model than 64-bit strong PUF, so the stability after screening is higher. When the initial stability is poor, the machine learning model works better, and the stability after screening is even higher than the initial stability. This phenomenon is related to the modeling principle of machine learning. When the output in the data set is too biased towards 0 or 1, the loss function of the machine learning model is easily trapped in local optimum and loses the optimization effect. When the stability is high, the S value for evaluating the stability is all 1, which leads to that the modeling accuracy cannot be further improved and the improvement effect is lost.

[0042] The response distribution graph corresponding to the initial stimulus set not screened by the strong PUF stable stimulus screening method based on machine learning of the application is shown in FIG. 4(a), and the response distribution graph corresponding to the stable stimulus set obtained after screening the initial stimulus set by the strong PUF stable stimulus screening method based on machine learning of the application is shown in FIG. 4(b). In FIG. 4(a) and FIG. 4(b), black elements represent responses of 1, and white elements represent responses of 0. Since the information entropy of a binary source contains the maximum amount of information when the probability is 50%, the more random the distribution of 0 / 1 in the response set, the better the output performance of the strong PUF. From FIG. 4(a) and FIG. 4(b), it can be seen that the screened stimulus does not affect the randomness of the strong PUF, and the proportion of 0 / 1 in the response set is still very close to 50%, and the distribution of the two colors is uniform, and the randomness is excellent.

[0043] The autocorrelation coefficient distribution diagram of 100000 responses corresponding to the initial stimulus set screened by the strong PUF stable stimulus screening method based on machine learning of the application is shown in Figure 5(a), and the autocorrelation coefficient distribution diagram of 100000 responses corresponding to the stable stimulus set obtained after screening the initial stimulus set by the strong PUF stable stimulus screening method based on machine learning of the application is shown in Figure 5(b). In Figures 5(a) and 5(b), the horizontal axis represents the response shift amount, and the vertical axis represents the autocorrelation coefficient of the current response. The autocorrelation function is used to represent how the correlation between any two values of a signal changes with the change of the interval, and it is used for uniqueness testing. A set of responses can be regarded as a discrete signal, and the stronger the uniqueness between each bit of the response, the weaker the correlation, and the smaller the autocorrelation function value. The correlation calculation of the responses generated by the one hundred thousand random stimuli and the screened stimuli is performed respectively, and it can be known from Figures 5(a) and 5(b) that the method of the application has little effect on the uniqueness of the strong PUF response, and the ACF is 0.0063 within 95% confidence interval, which is very close to the ideal value 0.

[0044] In summary, the strong PUF stable stimulus screening method based on machine learning is proposed to improve the stability of the response. The method uses machine learning to model the relationship between the stimulus and the corresponding stability, and calculates the stability value of any stimulus through the screening model. When the unknown stimulus set is input to the strong PUF, unstable stimuli can be discarded, and the stability of the screened stable stimulus set will be greatly improved. It can be seen from the experimental results that the strong PUF stable stimulus screening method based on machine learning can effectively improve the stability of various APUF-based strong PUFs, and the stronger the external interference, the more obvious the effect. Compared with the traditional error correction algorithm, there is no risk of information leakage, because the stability of a specific stimulus can be determined using the proposed method before testing. The results show that the method can be widely used to improve the stability of strong PUFs, and provides an effective solution for anti-interference of Internet of Things device authentication.

Claims

1. A machine learning based strong PUF stable challenge selection method, characterized in that The method comprises the following steps: Step 1: randomly generating n groups of stimuli for testing the stability of the strong PUF, n being an integer greater than or equal to 100000, repeatedly inputting each group of stimuli in the n groups of stimuli into the strong PUF m times, m being an integer greater than or equal to 100, obtaining m responses corresponding to each group of stimuli, if the m responses corresponding to a group of stimuli in the n groups of stimuli are all the same, the group of stimuli is a stable stimulus, the stability of the corresponding response is 1, otherwise the group of stimuli is an unstable stimulus, the stability of the corresponding response is 0, thereby obtaining the stability of the response corresponding to each group of stimuli in the n groups of stimuli; Step 2: defining a group of stimuli and the stability of the corresponding response as a CSP (Challenge-Stability Pair), and constructing a CSP with each group of stimuli and the stability of the corresponding response in the n groups of stimuli, thereby obtaining n groups of CSPs, and using the n groups of CSPs to construct a machine learning training data set; Step 3: obtaining a machine learning model, adjusting the number of parameters to be solved in the machine learning model to be the same as the number of bits of a group of stimuli of the strong PUF, each parameter to be solved corresponding to a bit of a group of stimuli, taking a group of stimuli in each CSP as the input of the machine learning model, and taking the stability of the response corresponding to the group of stimuli as the output of the machine learning model, training the machine learning model with the n groups of CSPs until the machine learning model converges, at this time the machine learning model is trained, and the trained machine learning model is a screening model, wherein in the process of training the machine learning model, an optimization algorithm is used to perform minimum value optimization on the loss function of the machine learning model, and when the loss function is minimum, it indicates that the machine learning model converges; Step 4: in the actual application of the strong PUF, the screening model is used to screen stable stimuli, and the specific process is as follows: S1: randomly generating multiple groups of stimuli of the strong PUF, and constructing an initial stimulus set with the multiple groups of stimuli, the multiple groups of stimuli in the initial stimulus set being randomly sorted; S2: taking the multiple groups of stimuli in the initial stimulus set as the input of the screening model in sequence, the output of the screening model being the stability of the corresponding response of each group of stimuli input into the screening model, if the output of the screening model is 1, it indicates that the group of stimuli input into the screening model is a stable stimulus, if the output of the screening model is 0, it indicates that the group of stimuli input into the screening model is an unstable stimulus, before all the multiple groups of stimuli in the initial stimulus set are input into the screening model or after all the multiple groups of stimuli in the initial stimulus set are input into the screening model, if the number of stable stimuli obtained meets the actual application requirement, the input is stopped, if the number of stable stimuli obtained still does not meet the actual application requirement after all the multiple groups of stimuli in the initial stimulus set are input into the screening model, the step S1 is returned, and the number of stable stimuli obtained meets the actual application requirement; Step 5: using all the stable stimuli determined in step 4 to construct a stable stimulus set, and using the stable stimulus set as the input of the strong PUF for identity information extraction.

2. The machine learning based strong PUF stable challenge selection method of claim 1, wherein The machine learning model is any one of a support vector machine (SVM) model, an artificial neural network (ANN) model, a convolutional neural network (CNN) model and a light gradient boosting machine (LightGBM) model.

3. The machine learning based strong PUF stable challenge selection method of claim 1, wherein The optimization algorithm is any one of a gradient descent method, a Newton method and a swarm intelligence algorithm.

Images