A neural network-based crystal oscillator PUF and a working method thereof

Abstract

The present application belongs to the field of integrated circuits, and particularly relates to a crystal oscillator PUF based on a neural network and a working method thereof. The present application acquires a frequency change curve of a crystal oscillator from power-on to stable oscillation, which contains frequency f, temperature T and power supply voltage VDD, converts the frequency change curve into a digital signal and outputs the digital signal to a neural network control module; then, through a trained MLP neural network model, the received digital signal is mapped into a corresponding feature vector, and the feature vector is transmitted to a key generation module for generating a key. The present application strips the environmental noise of the crystal oscillator caused by external influences in the starting process through the neural network, restores the original information determined only by the physical structure of the crystal, and generates a key by using the original information, thereby reducing the redundant error correction circuit and improving the reliability of the crystal oscillator PUF, and providing a possibility for the wide application of the crystal oscillator PUF.

Description

Technical Field

[0001] This invention belongs to the field of integrated circuits, specifically relating to a neural network-based crystal oscillator (PUF) and its operating method. Background Technology

[0002] In recent years, with the continuous development of artificial intelligence and integrated circuit technologies, intelligent electronic devices related to integrated circuits have been widely used in modern society, making hardware circuit cloning a major challenge. Due to random process variations in chip manufacturing, Physically Unclonable Functions (PUFs) embedded in hardware circuits have the characteristic of preventing cloning. Therefore, stimuli applied to two identical intelligent electronic devices will produce two different responses. PUFs have the advantages of low power consumption, non-cloning, and no reliance on fixed storage, enabling the generation of unique electronic fingerprints for devices, thereby achieving secure and reliable authentication.

[0003] Existing research on physically unclonable functions (PUFs) can be broadly categorized into two types: delay-based PUFs and memory-based PUFs. Delay-based PUFs generate unique responses based on the time difference of signal propagation, such as the Ring Oscillator PUF (ROPUF), the Configurable RO PUF (CRO PUF), and the Arbiter PUF. Memory-based PUFs, on the other hand, generate unique identifiers based on the random initialization state of the memory cells upon power-up. For example, the Static Random Access Memory (SRAM) PUF generates a unique response based on the stable state ("0" or "1") of the SRAM after power-up.

[0004] Crystal oscillators (XOs) utilize the inherent frequency dispersion of quartz wafers due to variations in their cut, thickness, electrode geometry, and packaging. They provide a robust, low-power, and invasive physical non-cloning function implementation path for CMOS chips without the need for additional masks or post-processing. However, while their frequency response is an order of magnitude less sensitive to temperature and power supply voltage than CMOS ring oscillators, residual temperature and voltage drift can still cause errors during key reconstruction in automotive-grade applications. This necessitates error correction circuitry, increasing area, power consumption, and latency, severely limiting their application in low-power MCUs (microcontroller units), passive RFID (radio frequency identification), and implantable medical chips.

[0005] In recent years, machine learning technology has provided a new path to solve the challenges of analog circuit manufacturing processes, voltage, and temperature. Neural networks, with their powerful nonlinear fitting and real-time decision-making capabilities, have demonstrated significant advantages in scenarios such as analog circuit error calibration and parameter optimization. However, current research has not yet applied neural networks to power supply element fans (PUFs) to reduce the impact of temperature and power supply voltage on their reliability and to decrease the bit error rate of PUFs. This has become a significant factor restricting the commercialization of PUFs. Summary of the Invention

[0006] To address the aforementioned problems and shortcomings, and to resolve the issues of existing crystal oscillator PUF startup processes being extremely sensitive to temperature and power supply voltage, resulting in poor reproducibility and high bit error rate, this invention provides a crystal oscillator PUF based on a neural network. This neural network eliminates the influence of varying temperature and power supply voltage, thereby reducing the bit error rate of the crystal oscillator PUF and improving its reproducibility and reliability while avoiding complex error correction circuits.

[0007] A neural network-based crystal oscillator (PUF) includes: a crystal oscillator, a state acquisition module, a neural network control module, and a key generation module.

[0008] The status acquisition module is used to record the frequency-temperature curve and frequency-voltage curve of the crystal oscillator from power-on to stable oscillation, including the crystal oscillator frequency f, temperature T and supply voltage VDD. The acquired data information is converted into digital signals and then transmitted to the neural network control module.

[0009] The neural network control module is a pre-trained MLP neural network model and is the core control unit of this invention. It receives the crystal oscillator frequency f, temperature T, and supply voltage VDD from the state acquisition module as input vectors. Through the "frequency-temperature-supply voltage-environmental noise-original information" mapping learned during offline training, it extracts environmental noise from the signal interfered with by temperature T and supply voltage VDD, restoring the original information determined solely by the crystal's physical structure. This original information is then output as a feature vector to the key generation module, reducing the bit error rate of the crystal oscillator PUF.

[0010] The key generation module receives feature vectors from the neural network control module and generates corresponding keys based on the obtained feature vectors.

[0011] Furthermore, the crystal oscillator includes two capacitors, one resistor, one crystal, and one inverter; the resistor, crystal, and inverter are connected in parallel, and one end of each of the two capacitors is connected to both ends of the resistor, while the other end is grounded.

[0012] Furthermore, the core part of the key generation module is to realize the conversion from feature vector to key through Toeplitz matrix multiplication. Let the input vector be S and the output key be K. The conversion formula is: K=Z*S, where Z is a Toeplitz matrix generated by a random seed.

[0013] Furthermore, the key generation module preprocesses the input vector S before use: first, the input vector S is quantized and converted into a binary stream; then, an extractor architecture is used, combined with a random salt value to eliminate the bias in the input vector S; finally, when the entropy extracted by the extraction unit is insufficient to be directly used as a long key, it is extended by the key derivation function KDF.

[0014] Furthermore, the above-mentioned working method of the neural network-based crystal oscillator PUF includes the following steps:

[0015] Step 1: The status acquisition module acquires the frequency change curve of the crystal oscillator circuit from power-on to stable oscillation, including frequency f, temperature T and supply voltage VDD, and converts it into a digital signal to be output to the neural network control module.

[0016] Step 2: After receiving the frequency f, temperature T, and power supply voltage VDD information from the state collection module, the neural network control module eliminates the influence of temperature and power supply voltage according to the mapping relationship in the offline trained MLP neural network model, removes environmental noise from the interference signal, restores the original information determined only by the crystal oscillator itself, and outputs the feature vector to the key generation module.

[0017] Step 3: After receiving the feature vector of the MLP neural network model, the key generation module generates the corresponding key.

[0018] In summary, the core innovation of this invention lies in replacing traditional redundant error correction circuits with neural networks. By eliminating the influence of temperature and power supply voltage through neural networks, the original information determined solely by the crystal's physical structure is restored, significantly improving the accuracy and reliability of the PUF (Programmable Array Function) of the crystal oscillator. Simultaneously, the adaptive capability of the neural network eliminates the need for additional calibration and error correction circuits, simplifying system design complexity. This invention achieves a low bit error rate for the PUF while maintaining low complexity and high robustness, providing an efficient solution for the application of PUFs in high-security, low-power scenarios. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the workflow of the present invention.

[0020] Figure 2 This is a structural diagram of an embodiment;

[0021] Figure 3 This is a schematic diagram of the MLP neural network model used in the embodiment. Detailed Implementation

[0022] The technical solution of the present invention will be described in more detail below with reference to the embodiments and accompanying drawings.

[0023] A type of neural network-based crystal oscillator PUF, such as Figure 2 As shown, it includes: a crystal oscillator, a state acquisition module, a neural network control module, and a key generation module.

[0024] The status acquisition module acquires the frequency change curve of the crystal oscillator from power-on to stable oscillation, including frequency f, temperature T, and supply voltage VDD, and converts it into a digital signal to be output to the neural network control module. After receiving the data from the status acquisition module, the neural network control module eliminates the influence of temperature and supply voltage according to the mapping relationship of the pre-trained MLP neural network model, removes environmental noise from the interference signal, restores the original information determined only by the crystal oscillator itself, and outputs a feature vector to the key generation module. The key generation module generates a key based on the feature vector transmitted from the neural network control module.

[0025] The detailed structure and working principle of each part are explained below:

[0026] Reference Figure 1 Here is a flowchart of the workflow of this invention, and the steps are as follows:

[0027] Step 1: The status acquisition module acquires the frequency change curve of the crystal oscillator from power-on to stable oscillation, including frequency f, temperature T and supply voltage VDD, and converts it into a digital signal to be output to the neural network control module.

[0028] Step 2: After receiving the frequency f, temperature T, and power supply voltage VDD information from the state collection module, the neural network control module outputs a feature vector to the key generation module according to the mapping relationship in the trained neural network.

[0029] Step 3: After receiving the feature vector from the neural network control module, the key generation module generates the corresponding key.

[0030] In this embodiment, the key generation module preprocesses the feature vector (i.e., input vector S) output in step 2 before use: First, the input vector S is quantized and converted into a binary stream; then, an extractor architecture is used, combined with a random salt value to eliminate bias in the input vector S; when the entropy extracted by the extraction unit is insufficient to be directly used as a long key, it is extended through the key derivation function KDF. Finally, the key generation module realizes the conversion from feature vector to key through Toeplitz matrix multiplication. Let the input vector be S and the output key be K, the conversion formula is: K=Z*S, where Z is a Toeplitz matrix generated by a random seed.

[0031] Reference Figure 3 This is a schematic diagram of the MLP neural network model used in this embodiment. It consists of an input layer, a hidden layer, and an output layer. Through the nonlinear mapping of multiple neurons, it realizes feature extraction of complex data. It has a simple structure, fast inference speed, and can quickly map frequency, temperature, and power supply voltage information to the corresponding feature vectors to achieve stable output.

[0032] The working and training mechanism of the MLP neural network model is as follows:

[0033] Training phase (Supervised Learning): Collect sample data of "temperature-power supply voltage-frequency" under different startup states of the crystal oscillator to construct a training dataset; with the goal of "extracting environmental noise from interference signals and restoring the original information determined only by the crystal oscillator itself", train the neural network to enable it to have the ability to accurately identify frequencies, adapt to PVT, and output stable feature vectors.

[0034] Inference phase: In actual operation, the MLP neural network model does not need to dynamically adjust the weights, but directly uses the trained fixed weights to output the feature vector, ensuring the stability of the feature vector and reducing the bit error rate of the crystal oscillator PUF.

[0035] As can be seen from the above embodiments, this invention uses a state acquisition module to collect the frequency change curve of the crystal oscillator from power-on to stable oscillation, including frequency f, temperature T, and supply voltage VDD, and converts it into digital signals for output to the neural network control module. Then, a trained MLP neural network model maps the received digital signals into corresponding feature vectors and transmits them to the key generation module for key generation. During training, the MLP neural network model has already collected "temperature-supply voltage-frequency" sample data of the crystal oscillator under different startup states to construct a training dataset, achieving accurate feature vector output. This invention uses a neural network to remove environmental noise generated by external influences during the crystal oscillator startup process, restoring the original information determined only by the crystal's physical structure, and uses this information to generate a key. This reduces redundant error correction circuits and improves the reliability of the crystal oscillator PUF, providing a possibility for the widespread application of crystal oscillator PUFs.

Claims

1. A neural network-based crystal oscillator (PUF), characterized in that, include: Crystal oscillator, status acquisition module, neural network control module, and key generation module; The state acquisition module is used to record the frequency-temperature curve and frequency-voltage curve of the crystal oscillator from power-on to stable oscillation, including the crystal oscillator frequency f, temperature T and supply voltage VDD. The acquired data information is converted into digital signals and transmitted to the neural network control module. The neural network control module is a pre-trained MLP neural network model. It receives the crystal oscillator frequency f, temperature T, and supply voltage VDD from the state acquisition module as input vectors. Through the frequency-temperature-supply voltage-environmental noise-original information mapping relationship learned in the offline training stage, it extracts the environmental noise from the signal interfered by temperature T and supply voltage VDD, restores the original information determined only by the crystal physical structure, and outputs it to the key generation module in the form of feature vectors. The key generation module receives feature vectors from the neural network control module and generates corresponding keys based on the obtained feature vectors.

2. The PUF crystal oscillator based on a neural network as described in claim 1, characterized in that: The crystal oscillator includes two capacitors, one resistor, one crystal, and one inverter; the resistor, crystal, and inverter are connected in parallel, and one end of each capacitor is connected to the two ends of the resistor, while the other end is grounded.

3. The PUF crystal oscillator based on a neural network as described in claim 1, characterized in that: The key generation module converts the feature vector into a key using Toeplitz matrix multiplication. Let the input vector be S and the output key be K. The conversion formula is: K = Z * S, where Z is a Toeplitz matrix generated from a random seed.

4. The neural network-based PUF crystal oscillator as described in claim 3, characterized in that: The key generation module preprocesses the input vector S before use: first, it quantizes the input vector S and converts it into a binary stream; then, it uses an extractor architecture and combines a random salt value to eliminate the bias in the input vector S; finally, when the entropy extracted by the extraction unit is insufficient to be used directly as a long key, it is extended by the key derivation function KDF.

5. The operating method of the neural network-based PUF crystal oscillator as described in any one of claims 1-4, characterized in that: Step 1: The status acquisition module acquires the frequency change curve of the crystal oscillator circuit from power-on to stable oscillation, including frequency f, temperature T and supply voltage VDD, and converts it into a digital signal to be output to the neural network control module. Step 2: After receiving the frequency f, temperature T, and power supply voltage VDD information from the state collection module, the neural network control module eliminates the influence of temperature and power supply voltage according to the mapping relationship in the offline trained MLP neural network model, removes environmental noise from the interference signal, restores the original information determined only by the crystal oscillator itself, and outputs the feature vector to the key generation module. Step 3: After receiving the feature vector of the MLP neural network model, the key generation module generates the corresponding key.

Images