Improved cryptanalysis combining differential and artificial neural network schemes

In this work we show the application of a neural cryptanalysis approach to S-DES input-output-key data to test if it is capable of mapping the relations among these elements. The results show that, even with a small amount of samples (about 0,8% of all data), the neural network was able to map the relation between inputs, keys and outputs and to obtain the correct values for the key bits k0, k1 and k4. By applying differential cryptanalysis techniques on the key space, it was possible to show that there is an explanation about the neural network partial success with some key bits. After implementing new s-boxes, which are more resistant to the differential attack, the neural network was not able to point out bits of the key any more. We believe that this new methodology of attack and repair assessment using neural networks has the potential to contribute in the future analysis of other cryptographic algorithms.