Research On Security Analysis Of The Lightweight Block Cipher Based On Deep Learning

The Internet of Things,big data systems,and various intelligent applications are ubiquitous in all aspects of people’s lives as the rapid development of science and technology,and a large amount of information is continuously generated.Subsequently,there are security issues with this information in an open network environment.Because of the same encryption and decryption structure,fast execution speed and easy standardization,the block ciphers are widely used in the field of information security.In particular,lightweight block cipher,which is widely used in resource constrained environments,along with many application scenarios with different properties,security analysis and application analysis of lightweight block cipher become more and more important.The classical cryptanalysis mainly adopts automated search technologies,such as mixed integer linear programming method.The technologies of deep learning have shown significant impact on various fields and have become current research hotspots.It is an urgent problem to be solved to build intelligent models for the distinguisher,the core component of cryptanalysis,and explore new tools for cryptanalysis in the development of block cipher intelligent analysis technology.This paper focuses on the security analysis of lightweight block cipher based on the deep learning method and the application of lightweight block cipher in source location information protection.The main contents include the construction of distinguisher based on deep learning for lightweight block cipher SIMON,SIMECK,ASCON with large block size and the Chinese national block cipher SM4,the key recovery of SIMON based on neural distinguisher and the application analysis of lightweight block cipher.The main research of the paper is as follows:（1）A method to train the deep learning distinguisher model based on differential cryptanalysis and ciphertext pair is proposed,and the deep learning distinguisher for the lightweight block cipher SIMON32/64 is constructed.Adopting the idea of differential cryptanalysis,two deep learning models,convolutional neural network and feedforward neural network,are used to build the neural distinguishers.The training of the neural distinguisher model uses one ciphertext pair with a single differential.The dataset is constructed to train and validate the model,and 6-9 rounds of neural distinguishers of SIMON32/64 are obtained.The data complexity,time complexity,and storage complexity of the obtained neural distinguishers are lower than those of traditional differential distinguishers.The neural distinguishers are used to perform 11 rounds of key filtering and key recovery attack on the SIMON32/64.The data complexity,time complexity,and storage complexity of the 11-round attack using neural distinguishers are 2^10.229 chosen-plaintexts,2^26.304 encryptions and 2^16.026 bytes respectively,which are lower than the results of differential attacks.（2）A method to train the model of deep learning distinguisher based on multiple ciphertext pairs using multiple differentials is proposed,and the multiple differentials distinguisher base on deep learning for the lightweight block cipher SIMECK32/64 is constructed.Adopted the idea of plaintext difference set and difference cryptanalysis,two deep learning models,convolutional neural network and feedforward neural network,are used to build the multiple differential neural distinguishers.The dataset based on multiple differentials is constructed to train and validate the model.The 6-11 rounds of neural distinguishers of SIMECK32/64 are obtained when the convolutional neural network is used to train.The 6-10 rounds of neural distinguishers of SIMECK32/64 are obtained when the feedforward neural network is used to train.The impact of the number of differentials on the accuracy of the distinguishers model is analysis.The data complexity,time complexity,and storage complexity of the neural distinguishers are lower than those of traditional differential distinguishers.The accuracy of the obtained neural distinguishers is higher than that of the neural distinguishers built by using single ciphertext pair with single differential and multiple ciphertext pairs with single differential.The neural distinguisher has high reliability,and the accuracy of filtering incorrect ciphertext using 7-9 rounds of neural distinguishers is99.61%,96.10%,and 77.34%respectively.（3）A method to construct the deep learning distinguisher models for ciphers with large block is proposed,and the deep learning distinguishers are constructed on ASCON,a large state lightweight block cipher,and SM4,a national block cipher.Based on the idea that ciphertext difference can improve the performance of the distinguisher,partial differences between ciphertext pairs are also used as part of the training data.A new input data format for the neural distinguisher is designed,and a residual neural network model is adopted to construct the neural distinguisher.The data preprocessing is performed on training datasets.The 1-4 rounds neural distinguisher for ASCON and the 4-9 rounds neural distinguisher for SM4 are obtained.The round of neural distinguisher of ASCON has been improved from 3to 4,and its accuracy is higher than existing neural distinguishers.The data complexity,time complexity,and storage complexity of SM4 neural distinguishers are much lower than those of traditional differential distinguishers,and the accuracy of SM4 neural distinguishers is higher than existing neural distinguishers.The experimental results show that the deep learning is effective and feasible in the security analysis of block cipher with large block.（4）A method to protect the location information of the source node in wireless sensor networks based on lightweight block cipher is proposed.A multiple branch paths routing protocol EBMB is constructed based on virtual source nodes and virtual data.The software implementation of the lightweight block cipher SIMECK,SIMON,SPECK and ASCON etc.on the specific hardware platform are carried out,and the current main performance indicators of the above cryptographic algorithms are evaluated.Based on the analysis of the security of the lightweight block cipher algorithm by using the deep learning distinguishers,the hop-by-hop encryption and decryption mechanism in the source location information protection is implemented adopting SIMECK cryptographic algorithm.The experimental results show that,compared to the comparison method,the EBMB method proposed in this paper has a higher network security cycle and network lifetime.The improved method of hop-by-hop encryption and decryption based on SIMECK achieves data confidentiality while maintaining low network energy consumption.The proposed method has good data confidentiality and location information protection capabilities.