Representation vs. Architecture in Neural Differential Distinguishers

Neural differential distinguishers are widely used to evaluate round-reduced symmetric primitives, yet it remains unclear whether performance improvements arise primarily from larger neural models or from more informative input representations. We present a systematic study of this trade-off using ciphertext-only features under a chosen-plaintext differential sampling model on four lightweight block ciphers: the SPN ciphers GIFT-64, RECTANGLE-64, and PRESENT-64, together with the ARX cipher SPECK-64/128. We compare a standard single-difference representation, denoted R1, with a multi-difference matrix representation, denoted R2, and evaluate three common model families, namely MLP, CNN, and ResNet, under a fixed and uniform training protocol. Our results show that, in this setting, representation choice is a more consistent driver of performance than architecture choice. Across the main experiments, R2 generally yields stronger distinguishers than R1, whereas deeper models often provide only limited additional gains and can overfit in harder regimes. This conclusion is reinforced by supplementary experiments showing that the representation gap remains visible under an alternative shared difference set, persists for a simple linear baseline, and often survives under an equal-ciphertext budget. The results support the view that, for neural distinguishers in this setting, designing information-dense ciphertext representations can matter at least as much as scaling model capacity, and often more.