An Improved Integral Distinguisher Scheme Based on Deep Learning

EasyChair Preprint 4735

Abstract

At Crypto 2019, A.Gohr made a breakthrough in the fusion of differental cryptoanalysis and deep learning on Round Reduced Speck. Inspired by his methodology, we apply deep learning to construct a neural-based integral distinguisher scheme. To guage the advantage of our distinguisher scheme, we apply it on several block ciphers including SPECK32/64, SIMECK32/64, PRESENT,RECTANGLE and LBLOCK and compare it with bit-based division property as the state-of-the art integral distinguishing method. To our great surprise, our neural network based integral distinguisher extends the number of distinguished rounds of most block ciphers by at least 1 additional rounds compare to bit-based division property. In some cases, such as PRESENT block cipher, we could achive a 8-round distinguisher, while, bit-based division property could only provide a 5 round distinguisher. ‘ese observations illustrates that the fusion of deep learning and integral cryptanalysis has a promising prospect. In addition, we apply different deep learning architectures, namely ResNet,ResNeXt and DenseNet . DenseNet, as state of the art architecture, slightly outperforms the other architectures in terms of the distinguishing accuracy. Further, we showcase the utility of FewShot learning in reduction of training dataset to less than 100 instances. Finally, we show that our neural distiguisher is not only helpful for block cipher designers, but also assists a‹ackers to mount key recovery attack. To this end, we show how to exploit our distinguisher to mount key recovery attack and apply it to SPECK32/64.

Keyphrases: Differential cryptoanalysis, Division property, Integral cryptoanalysis, LBlock, PRESENT, SIMECK, SPECK, deep learning, rectangle

@booklet{EasyChair:4735,
  author    = {Behnam Zahednejad and Jin Li},
  title     = {An Improved Integral Distinguisher Scheme Based on Deep Learning},
  howpublished = {EasyChair Preprint 4735},
  year      = {EasyChair, 2020}}