Deep-Learning-Based Neural Distinguisher for Format-Preserving Encryption Schemes FF1 and FF3

Abstract

Distinguishing data that satisfy the differential characteristic from random data is called a distinguisher attack. At CRYPTO’19, Gohr presented the first deep-learning-based distinguisher for round-reduced SPECK. Building upon Gohr’s work, various works have been conducted. Among many other works, we propose the first neural distinguisher using single and multiple differences for format-preserving encryption (FPE) schemes FF1 and FF3. We harnessed the differential characteristics used in FF1 and FF3 classical distinguishers. They used SKINNY as the inner encryption algorithm for FF3. On the other hand, we employ the standard FF1 and FF3 implementations with AES encryption (which may be more robust). This work utilizes the differentials employed in FF1 and FF3 classical distinguishers. In short, when using a single 0x0F (resp. 0x08) differential, we achieve the highest accuracy of 0.85 (resp. 0.98) for FF1 (resp. FF3) in the 10-round (resp. 8-round) number domain. In the lowercase domain, due to an increased number of plaintext and ciphertext combinations, we can distinguish with the highest accuracy of 0.52 (resp. 0.55) for FF1 (resp. FF3) in a maximum of 2 rounds. Furthermore, we present an advanced neural distinguisher designed with multiple differentials for FF1 and FF3. With this sophisticated model, we still demonstrate valid accuracy in guessing the input difference used for encryption.