Key recovery attack on PRESENT using an entropy-based neural distinguisher

In 2019, Gohr introduced neural-distinguishers as a tool to improve differential cryptanalysis using deep learning. Building on this, we propose an Entropy-based Neural Distinguisher for present that requires significantly fewer input bits and network parameters. Our method identifies relevant bit subsets by analyzing the entropy of output differences between ciphertext pairs. We reduce the distinguisher size by: (i) simplifying the network architecture, and (ii) shrinking the input layer via entropy-based bit selection. Our distinguisher achieves accuracy within 1% (resp. 4%) of the state of the art for 7 (resp. 6) rounds of present, using only 28 out of 64 bits and under 10% of the original parameters. Leveraging this efficient setup, we introduce an iterative key recovery method capable of handling 64-bit round keys—unlike Gohr’s 16-bit target. Our approach recovers full keys (64 bits) with 92.8% success (402 out of 433), with all partial recoveries retrieving at least 56 bits and 83.8% retrieving 60 bits or more.