Do AES encryptions act randomly?

The Advanced Encryption Standard (AES) is widely recognized as the most important block cipher in common use nowadays. This high assurance in AES is given by its resistance to ten years of extensive cryptanalysis, that has shown no weakness, not even any deviation from the statistical behaviour expected from a random permutation. Only reduced versions of the ciphers have been broken, but they are not usually implemented. In this paper we build a distinguishing attack on the AES, exploiting the properties of a novel cipher embedding. With our attack we give some statistical evidence that the set of AES-$128$ encryptions acts on the message space in a way significantly different than that of the set of random permutations acting on the same space. While we feel that more computational experiments by independent third parties are needed in order to validate our statistical results, we show that the non-random behaviour is the same as we would predict using the property of our embedding. Indeed, the embedding lowers the nonlinearity of the AES rounds and therefore the AES encryptions tend, on average, to keep low the rank of low-rank matrices constructed in the large space. Our attack needs $2^{23}$ plaintext-ciphertext pairs and costs the equivalent of $2^{48}$ encryptions. We expect our attack to work also for AES-$192$ and AES-$256$, as confirmed by preliminary experiments.