Plaintext restoration of light weight block ciphers using deep neural networks under the black box assumption

Classical cryptanalytic attacks i.e., linear and differential attacks are used for evaluating the security strength of symmetric-key based block ciphers, which are known to be the power house of digital security. These attacks exploit underlying weaknesses in the internal structure of the targeted cipher and hence, are highly specific to the confusion-diffusion paradigm of the targeted ciphers. However, high computational complexity of classical cryptanalytic techniques limits their effectiveness, even against the reduce-round variants of the targeted ciphers. In view of the limitations of conventional cryptanalysis and recent successful implementation of deep-learning based neural distinguisher against Speck-32/64 (CRYPTO-2019) in a black box setting, where attacker merely knows the algorithm interfaces, cryptanalysts were enticed to explore the viability and efficacy of deep-learning based cryptanalysis. Although several neural models have been proposed for plaintext restoration of DES, 3-DES, AES, Blowfish and other Light Weight Cryptographic (LWC) symmetric ciphers, yet, it has been observed that the computational complexity and number of rounds attacked by proposed neural model rounds are either equal or lesser than those attacked by traditional cryptanalysis. This necessitates the development of an optimized deep neural model for recovering plaintexts of a symmetric-key block ciphers in a known-plaintext threat scenario under a black-box assumption. Thereby, in this work Long Short-Term Memory (LSTM) based efficient neural cryptanalytic models have been proposed against two LWCs namely substitution permutation networked PRESENT-64/80 (previously attacked till round-3 with a success probability of 2^–3.14) and Feistel structured SIMON-64/96; which till date has not been subjected to any neural Plaintext Recovery (PR) attack earlier. Moreover, efficiency of the proposed neural models has been verified by evaluating hamming-distance based accuracy against seven different datasets of varying correlation complexities. These datasets vary from Correlated Train Test (CTT), Non-Correlated Train Test (NCTT) to Real-World Train Test (RWTT) ciphertext-plaintext samples. Based on the experimental results, it is observed that complete encryption oracle of PRESENT and SIMON light weight ciphers can be restored within an accuracy range of 54.12–69.80% and 55.11–68.3%, respectively, which to the best of authors knowledge, are the only and highest neural plaintext restoration accuracies achieved against the full rounds of the targeted LWCs. Furthermore, the accuracy ranges observed against different confusion-diffusion paradigms of the targeted LWCs also reaffirms that the proposed neural PR attack is independent of the internal architecture of the targeted cipher and thus, has a potential to be generalized against different variants of symmetric block ciphers.