High-Accuracy Min-Entropy Assessment Method Based On TPA-LSTM Prediction Model

Random numbers are widely used in the field of information security,such as key generation and electronic signatures,and are the security guarantee of encryption systems.Safe random numbers are unpredictable,and their randomness is mainly derived from entropy sources.In practical applications,entropy sources should not be considered safe by default.External attacks,noise,and the presence of non-ideal characteristics of the physical devices can compromise the security of the entropy source and thus undermine the unpredictability of the random number,the security assessment of the entropy source is important.At present,the safety assessment methods of entropy sources mainly include statistical methods and entropy assessment methods.Statistical methods are to assess the safety of entropy sources by detecting whether the "0" and "1" binary sequences produced by entropy sources meet specific hypothesis test criteria.But statistical methods are difficult to detect hidden relationships between data;For example,some pseudo-random entropy sources with longer periods are able to pass statistical methods.The entropy assessment methods use the entropy value to measure the safety of the entropy source,which mainly include Shannon entropy,permutation entropy,min-entropy.Among them,min-entropy is a method proposed by the National Institute of Standards and Technology in SP 800-90 B standard to evaluate the safety of entropy sources.Min-entropy has been widely used in the field of information security,which uses machine learning algorithms to mine the hidden relationship between entropy source data.However,due to the lack of generalization and data processing power of machine learning algorithms,90 B has the problem of overestimation or underestimation in the measurement of min-entropy,and cannot handle large-scale data.In recent years,long-term and short-term memory neural networks and attention mechanisms has shown strong generalization capabilities and data processing capabilities,which is expected to solve the shortcomings of machine learning in the assessment of entropy source security.In this thesis,a min-entropy assessment method based on long-short-term memory neural network and attention mechanism(TPA-LSTM)prediction model is proposed.The TPA-LSTM prediction model is able to learn the long-term dependencies and inherent characteristics that may exist in the time series,thereby improving the accuracy of the min-entropy assessment.In addition,the safety of the physical entropy source was evaluated by the above method,and the influence of various parameters in the physical entropy source on its performance was studied,and the optimal working parameters were determined.The specific research content of this thesis is as follows:1.A min-entropy assessment method based on TPA-LSTM prediction model is proposed.First,the TPA-LSTM prediction model is designed,including a one-hot encoder,a long shortterm memory layer,an attention mechanism layer,and a fully connected layer.Next,the entropy source sample data is standardized,labeled,and divided into training set,validation set,and test set.The optimization of weights and hyperparameters in the TPA-LSTM prediction model is achieved by using the training set and the validation set,which improves the accuracy of the min-entropy assessment.Finally,the test set is used to verify the effectiveness of the minentropy assessment method proposed in this thesis.2.The accuracy of min-entropy assessment based on TPA-LSTM prediction model is explored.The accuracy of the TPA-LSTM predictor,the machine learning predictor,the FNN predictor,and the RNN predictor were evaluated.The above predictors were used to detect the random datasets of four known theoretical min-entropy and the relative error between the detection value and the theoretical value is calculated.Experimental results show that the relative error of TPA-LSTM predictor can reach 0.34%,which is smaller than the relative error of other types of predictors,which proves that the TPA-LSTM prediction model can more accurately assess the safety of entropy sources.3.The safety of white chaos physical entropy sources was evaluated using the TPA-LSTM predictor.First,two chaotic lasers are generated by using an external cavity semiconductor laser,and a white chaos with a bandwidth of 15.8 GHz is generated by optical heterodyne.Subsequently,the evaluation of min-entropy based on the TPA-LSTM prediction model proves that optical heterodyne can improve the min-entropy of the entropy source.Finally,the effect of classical noise on the performance of the white chaotic entropy source in the experimental apparatus is explored,and the interference of noise on the minimum entropy detection is eliminated by intercepting the effective bits.In addition,the effect of feedback intensity,current,and wavelength detuning on the safety of white chaotic entropy sources was evaluated using the TPA-LSTM predictor.Experimental results show that the min-entropy of the white chaos entropy source reaches the maximum value when the feedback intensity is-15 d B,the current is 28 m A,and the wavelength detuning is 0.15 nm.The output of the white chaotic entropy source under different parameters indicating that the delay feedback characteristics of the outer cavity semiconductor lasers were the reason for the poor safety.