Research On Automatic Analysis Methods Of Differential Classification In Block Ciphers

Block cipher plays an important role in the area of the networks and information security because of its advantages such as fast encryption and decryption,easy standardization,easy evaluation of its security,and so on.Differential cryptanalysis of block ciphers is a classical analysis method whose basic idea is to look for differential characteristics with the highest probability（differential distinguisher）,and then the key recovery attack is mounted.The automatic construction of the new differential distinguishers in block cipher appears to be a current research hotspot.In this thesis,some differential neural distinguisher automated search methods are proposed by basing on deep learning.Moreover,new security evaluation for round-reduced SPECK32/64,SKINNY-64,and PRESENT-80 encryption algorithms are investigated.The main research results are given as follows:1.A new search method for multiple differential neural distinguishers is proposed.The method is based on the idea of multiple differential cryptanalysis.In the first place,the differential characteristics is determined by MILP（Mixed Integer Linear Programming）technique,which is combined with the Res Net structure in deep learning to construct a multiple differential neural distinguisher.Moreover,the security of SPECK32/64 encryption algorithm is checked via the search method above.It illustrates that the differential neural distinguishers of SPECK32/64 are captured for 5-round to 7-round,(all with a data complexity of 2^24.99).Compared with previous known results,the method increases the accuracy of the 5-round distinguisher by 5%and the accuracy of the 6-round distinguisher by 7.6%.2.A new automatic search method is presented for the 5-round（or 6-round）differential distinguisher of SKINNY-64 cipher.Both the structure and parameters of the single differential（or multiple differential）neural distinguisher models are adjusted via the characteristics of the SKINNY-64 cipher.More precisely,the 5-round and 6-round differential characteristics of the SKINNY-64 cipher are respectively selected by the MILP technique,and then the neural distinguisher of the 5-round and 6-round are respectively achieved according to its input difference.The experimental results show that the data complexity of single differential and multiple differential models are all 2^23.39 and 2^24.99,respectively,which are significantly reduced compared to the traditional distinguisher(data complexity of 2⁴⁵).In particular,the accuracy of the model for the 5-round SKINNY-64differential distinguisher is about 99.9%,and about 75%for the 6-round distinguisher.3.A new automatic search method for the differential distinguisher of PRESENT-80cipher is described.More concretely,the novel multiple differential neural distinguisher model is constructed by combining both the MLP（Multi-Layer perceptron）structure in deep learning and the multiple differential.Compared with the existing results,the MLP structured multiple differential models extends the number of neural distinguisher rounds of the PRESENT-80 cipher from the original 5-round to 7-round,and the data complexity of the 5-round to 7-round differential neural distinguisher are all only about 2^24.99 choices of plaintext.