A method for expanding a random telegraph noise signal based on a memory neural network

Abstract

The application discloses a method for expanding random telegraph noise (RTN) based on a storage neural network, and the signal expansion process is realized based on an artificial neural network of a novel storage unit. According to partial RTN measured signals as a learning set, the expansion process of signal prediction reasoning can realize expansion of the signals with an arbitrary time length, and accelerates extraction of the time parameters of the RTN. The method has important significance for development of a physical unclonable function (PUF) technology based on the RTN and information data security.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor devices, and particularly relates to a method for expanding random telegraph noise signals based on an artificial neural network. BACKGROUND

[0002] In recent years, the use of mobile and embedded devices in daily work has steadily increased, and network security, hardware security, and information security are of vital importance to individuals, enterprises, and society. In response to hardware security and attack-resistant design, for traditional technologies, the generation of secret keys increases the cost and complexity of the manufacturing process, while limiting flexibility; at the same time, the storage mode of secret keys also faces many problems, such as secret keys stored in non-volatile memory being vulnerable to hardware attacks, and increasing the cost and manufacturing complexity of adding security elements. Emerging technologies are based on inherent random keys in chips, providing protection for hardware intrinsic security (HIS) systems. Physical unclonable functions (PUF) use physical layer random distribution as an entropy source, and for each input (C), a specific unique output (P) is generated. Therefore, for each chip, the input response pair (CRP) of the PUF is unique, thereby making the chip have anti-counterfeiting characteristics. PUF has three characteristics, namely, uniqueness: different PUFs return different responses to the same input; reliability: the same PUF returns the same response to the same input; unpredictability: the returned value of the PUF response cannot be predicted in advance. PUF is an effective and feasible method that not only reduces the cost and complexity of manufacturing, but also improves security. The performance of PUF is related to the adoption of entropy sources and architecture design, among which, an ideal entropy source usually plays a dominant role.

[0003] In semiconductor device manufacturing, the number of interface state traps and oxide layer defects decreases with the miniaturization of device size, and when the traps / defects are reduced to several orders of magnitude, the random behavior of carrier capture and release causes the random telegraph noise (RTN) phenomenon. At present, many new PUFs based on different entropy sources have appeared. Random telegraph noise (RTN), which originates from the random fluctuations of physical layer current or voltage, is a good choice for designing PUF random entropy sources. The random distribution of defects caused by devices during the manufacturing process will produce specific random telegraph noise signals for a specific device, making different devices have different RTN responses. Therefore, the RTN signal has unpredictability and uniqueness. Most PUFs use device threshold voltage (V th ) variations, such as current mirror (SRAM) PUF, ring oscillator (RO) PUF, etc. However, the distribution of V th follows a Gaussian distribution, and V thThe time drift may occur due to the persistent voltage stress, and the degradation of the device destroys the chip fingerprint. Therefore, it is beneficial to improve the uniqueness and reliability of the PUF by using the RTN phenomenon based on the physical layer randomness as a stable and durable entropy source and uniformly distributing the RTN time constant on a logarithmic time scale.

[0004] The RTN time constant is one of the representative device intrinsic parameters, and the capture time constant The emission time constant is defined as the average time required for the defect to release the carrier in the capture state. The emission time constant is defined as the average time required for the defect to release the carrier in the capture state. The accuracy of the RTN time constant is closely related to the test precision, sampling time, extraction method, etc. In addition, the length of the RTN measured signal has a significant impact on the extraction of the RTN time constant. When extracting the time constant based on the hidden Markov model (HMM), the longer the signal length, the higher the accuracy. Usually limited by test conditions or device characteristics, the length of the RTN measured signal is short, which is easy to cause errors. Therefore, the prediction expansion based on a small amount of RTN signals is realized, the signal length is effectively prolonged, and the time constant is quickly extracted. It has important research value and application significance for developing a new type of PUF based on RTN. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a method for expanding a random telegraph noise signal based on a storage neural network, which realizes the expansion of the RTN signal to any length, and further accelerates the extraction of the RTN time constant.

[0006] To solve the technical problem, the technical scheme adopted by the present application is as follows: a method for expanding a random telegraph noise signal based on a storage neural network, comprising the following steps:

[0007] S01), constructing a storage neural network, the storage neural network comprising a mask layer, a reservoir pool and an output layer, the mask layer being connected to the input, the output layer being connected to the output, the reservoir pool being located between the mask layer and the output layer, and the mask layer, the reservoir pool and the output layer being connected in a full connection or one-to-one connection manner;

[0008] S02), acquiring a random telegraph noise signal, performing binaryzation processing on the random telegraph noise signal to form a digital random telegraph noise signal, further performing a base change on the digital random telegraph noise signal to form an N-base random telegraph noise signal suitable for the storage neural network, N being a power of 2;

[0009] S03), inputting the N-base random telegraph noise signal as a learning set into the storage neural network, and training the N-base random telegraph noise signal;

[0010] S04), after the training, the first M-1 bits of the N-ary signal are inputted into the reserve pool of the storage neural network as preset, and the Mth bit of the N-ary signal is inputted into the storage neural network, so that an N-ary extended signal is obtained from the output layer, and the expansion of one bit of the random telegraph noise signal is completed, and M is a positive integer;

[0011] S05), the expanded signal is inputted into the storage neural network as a new input signal, so that the next bit of the expanded signal is obtained, and the expansion process is repeated to obtain more expanded information.

[0012] S-6), after the length of the expected expansion is completed, the N-ary signal is restored to a binary signal, and the expanded random telegraph noise signal is obtained.

[0013] Further, the storage neural network is a convolutional neural network, a recurrent neural network, a long short-term memory network or a reserve pool network formed by one of a resistive random access memory, a phase change memory, a magnetic resistive random access memory, a ferroelectric tunnel junction, a ferroelectric capacitor and a ferroelectric transistor.

[0014] Further, the storage neural network is a reserve pool network composed of a ferroelectric transistor, the mask layer and the reserve pool are connected in a full connection mode, and the reserve layer and the output layer are connected in a one-to-one connection mode.

[0015] Further, N is equal to 4, 8, 16 or 32.

[0016] Further, the method is suitable for the expansion of the Mackey-Glass time series and the Eronen mapping.

[0017] Further, the training in step S03) is to modulate the output layer to the desired optimized state, and the mask layer and the reserve pool do not need to be trained.

[0018] Further, the process of extracting the time constant of the expanded random telegraph noise signal is that, for a given length of random telegraph noise signal, the average time required for a defect to capture a carrier in a release state is the capture time constant , and the average time required for a defect to release a carrier in a capture state is the emission time constant .

[0019] The present application has the following advantages: the storage neural network is used to expand the random telegraph noise signal, a small amount of random telegraph noise (RTN) signal can be expanded to any length, and the extraction of the time constant is accelerated, and the extracted time constant is used as the entropy source of the PUF. The storage neural network can be trained to expand a small amount of RTN signal to any length, and the extraction of the RTN time parameter is accelerated, which is of great significance for the development of the physical unclonable function (PUF) technology based on the RTN and the information data security. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of matching and verifying PUF based on RTN;

[0021] Figure 2 This is a schematic diagram of a fully ferroelectric transistor reservoir network structure;

[0022] Figure 3 This is a flowchart of the RTN signal extension as a PUF processing method;

[0023] Figure 4 This is a comparison chart of the extended signal and the target signal after RTN is processed using an all-ferroelectric transistor reservoir network.

[0024] Figure 5 The result is the extraction of time constants based on a Hidden Markov Model (HMM) after processing RTN using an all-ferroelectric transistor reservoir network.

[0025] Figure 6 It is a bitmap of the extended signal and the target signal after processing RTN using an all-ferroelectric transistor reservoir network. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0027] Example 1

[0028] As device dimensions shrink, random telegraph noise becomes a major source of instability in nanoscale devices and circuits. During manufacturing, defects are unavoidable and cannot be replicated. Traps in the gate and substrate alternately trap and emit charge carriers. This charge migration perturbation-induced random telegraph noise manifests in the device as discrete oscillations of the transistor channel current and perturbations of the gate voltage. When charge is trapped, the current increases, and V... th Decrease; when charge is emitted, the current decreases, V th Increase. Due to the randomness of physical defect behavior, RTN becomes a good source of entropy. This property of the RTN time constant satisfies the PUF's requirements for uniqueness, reliability, and unpredictability.

[0029] For the RTN time constant, the average time required for a defect to capture carriers in the released state is defined as the capture time constant. The average time required for the defect to release carriers in the captured state is equal to the emission time constant. . and The average values of the capture time and the emission time correspond to the capture time and the emission time respectively, and together represent the difficulty of the defect to capture and release carriers, and provide information about the location and energy level of the trap. In the example of using RTN as a PUF entropy source, the time constant is defined as the case where the trap energy and the Fermi level are aligned, and in the device, It can be expressed as a uniform distribution in the logarithmic time. Thus, in the PUF, the is used as a parameter, which ensures its good randomness and robustness.

[0030] For the calculation of the above-mentioned RTN time constant, a time length suitable RTN is needed to count the time of carrier capture and release, and the value of the time constant is obtained according to the average value. A sufficient number of RTN signals helps to improve the accuracy of RTN time signal extraction, so it is of practical significance to expand the limited RTN signal. A method for expanding random telegraph noise signal based on storage neural network is proposed in this patent, which realizes the expansion of a small amount of RTN signal to any length. After training, the artificial neural network has the ability to predict, so for RTN, the expansion is to find the characteristics and then predict the future signal.

[0031] This embodiment takes the RTN signal expansion method realized by the all-ferroelectric transistor reservoir network as an example to illustrate the effectiveness of the proposed method. The ferroelectric transistors all use Hf 0.5 Zr 0.5 O2 ferroelectric film, and the structure of the three layers of the mask layer, the reservoir and the readout layer is composed of ferroelectric transistors, the mask layer is fully connected with the reservoir, and the reservoir is one-to-one connected with the readout layer, as shown in Figure 2 The expansion process of the RTN signal is as follows:

[0032] Based on the convenience of the process and the effectiveness of the subsequent expansion, first, the signal is binarized into a digital signal, and the digital RTN signal is further processed into a hexadecimal signal, and the expansion process is the processing of the hexadecimal signal. Second, in order to obtain a reserve pool (RC) network with RTN signal expansion capability, a measured RTN signal is used as a learning set to train the RC network, and the output layer ferroelectric transistor is modulated to the desired optimized state, while the mask layer and the ferroelectric transistor device in the reserve pool do not need to be modulated, greatly simplifying the training process. Third, after training, the first four bits of the latest 5-bit hexadecimal signal are preset as input to the all-ferroelectric transistor reserve pool network, and the fifth hexadecimal signal is input into the network to obtain a 1-bit hexadecimal expansion signal from the output, completing the expansion of one bit of the RTN signal. In order to obtain more expansion information of the RTN signal, the expanded signal is transmitted as a new input signal into the RC network for processing to obtain the next bit of the expanded signal. Finally, after the desired expansion length is completed, the hexadecimal signal is restored to a binary signal to obtain the expanded RTN signal. In this way, the RTN signal is expanded without length limitation, and the process is as shown in Figure 3 .

[0033] The effectiveness of the above-mentioned all-ferroelectric transistor reserve pool network for RTN signal expansion method is verified as follows. A 0.6s RTN signal is expanded by the above-mentioned method, as shown in Figure 4 . The gray color is the test signal, and the black color is the expanded binary signal. The 0.1s expanded signal and the target signal are displayed, and both have similar time constants. For different expansion ratios, the time constant extraction results are as shown in Figure 5 . When extracting the time constant of the random telegraph, based on the hidden Markov model (HMM), for a given length of RTN signal, the duration of charge capture / release is counted, and then processed and analyzed to obtain / . The expansion ratio is defined as the ratio of the expanded signal to the training signal, and the prediction error increases with the increase of the expansion ratio. The above verification proves that the all-ferroelectric transistor reserve pool network can quickly expand the RTN, and the expanded RTN signal ensures the effectiveness of the time constant. A 32x32 bit image is used to represent the target signal and the expanded signal, as shown in Figure 6 . The white color corresponds to charge capture, and the black color represents charge emission. The random distribution pattern indicates that the measured signal and the expanded signal have similar time constants.

[0034] The method has the expansion capability for slow RTN traps and fast RTN traps, and is not limited to the expansion of RTN signals, and has good expansion capability for other space-time signals such as Mackey-Glass time series and Hénon map.

[0035] In other embodiments, convolutional neural networks, recurrent neural networks, long short-term memory networks, and reservoir networks composed of new memory devices such as resistive random access memory, phase change memory, magnetic resistive random access memory, ferroelectric tunnel junctions, ferroelectric capacitors, and ferroelectric transistors can also be used.

[0036] In this embodiment, the binary signal is converted into a 16-bit signal for expansion, and according to the performance of the network device, the signal processed can be in the form of counting in four, eight, thirty-two, etc.

[0037] The above describes only the basic principles and preferred embodiments of the present application, and improvements and substitutions made by those skilled in the art based on the present application are within the scope of protection of the present application.

Claims

1. A method for extending random telegraph noise signals based on a storage neural network, characterized in that: Includes the following steps: S01) Construct a storage neural network, which includes a mask layer, a reservoir, and an output layer. The mask layer is connected to the input, the output layer is connected to the output, and the reservoir is located between the mask layer and the output layer. The mask layer, reservoir, and output layer are connected either fully or one-to-one. S02) Obtain the random telegraph noise signal, perform binarization on it to form a digital random telegraph noise signal, and further perform base transformation on the digital random telegraph noise signal to form an N-ary digital random telegraph noise signal adapted to the storage neural network, where N is a power of 2. S03) The N-ary digital random telegraph noise signal is used as a learning set to be input into the neural network for training. S04) After training, the first M-1 bits of the M-bit N-ary digital random telegraph noise signal are used as preset inputs to the storage pool of the storage neural network. Then, the M-th bit of the N-ary digital random telegraph noise signal is input to the storage neural network. One bit of the N-ary extended signal is obtained from the output layer, thus completing the extension of one bit of the N-ary digital random telegraph noise signal. M is a positive integer. S05) The extended signal is used as a new input signal and input into the storage neural network to obtain the next extended signal. The extension process is repeated to obtain more extended signals. S06) After completing the desired extension length, the N-ary digital random telegraph noise signal is restored back to a binary signal to obtain the extended digital random telegraph noise signal.

2. The method for extending random telegraph noise signals based on a storage neural network according to claim 1, characterized in that: The storage neural network is a convolutional neural network, recurrent neural network, long short-term memory network, or reservoir network formed by one of the following: resistive random access memory, phase change memory, magnetic resistive random access memory, ferroelectric tunnel junction, ferroelectric capacitor, and ferroelectric transistor.

3. The method for extending random telegraph noise signals based on a storage neural network according to claim 1, characterized in that: The storage neural network is a reservoir network composed of ferroelectric transistors. The mask layer and the reservoir are fully connected, and the reservoir layer and the output layer are connected one-to-one.

4. The method for extending random telegraph noise signals based on a storage neural network according to claim 1, characterized in that: N can be 4, 8, 16, or 32.

5. The method for extending random telegraph noise signals based on a storage neural network according to claim 1, characterized in that: This method is applicable to MacGregorian time series and extensions of the Eunon map.

6. The method for extending random telegraph noise signals based on a storage neural network according to claim 1, characterized in that: Step S03) The training involves modulating the output layer to the desired optimized state; the mask layer and the reservoir do not require training.

7. The method for extending random telegraph noise signals based on a storage neural network according to claim 1, characterized in that: The process of extracting the time constant from the extended digital random telegraph noise signal is as follows: For a given length of digital random telegraph noise signal, the average time required for the statistical defects to capture charge carriers in the released state is the capture time constant. The average time required for the defect to release charge carriers in the captured state is equal to the emission time constant. .

Images