Paper 2026/1530

Rich Input Representations in Neural Differential Cryptanalysis: A Taxonomy and Survey

Abstract

Neural differential distinguishers have become an active research direction in​ symmetric-key cryptanalysis since the introduction of deep-learning-based attacks on​ round-reduced SPECK. Early neural distinguishers typically used a single ciphertext pair​ or ciphertext difference as input. Recent studies, however, show that richer input​ representations can substantially affect the information available to the classifier, the data​ cost of each labeled sample, and the relevance of the distinguisher to practical attacks.​ Examples include multi-pair, multi-difference, matrix-style, multi-round,​ structured-encoding, and score-aggregation based inputs.​ This paper provides a taxonomy and survey of rich input representations in neural​ differential cryptanalysis. We introduce a representation-centric framework that describes​ an input representation by its difference set, number of observations per sample, sharing​ structure, encoding function, and ciphertext cost. Using this framework, we organize​ existing works into representation families and compare their motivations, benefits, and​ limitations. We also argue that representation-rich distinguishers require cost-aware​ evaluation: fixed-sample comparisons and fixed-ciphertext comparisons answer different​ questions and may lead to different conclusions. Finally, we identify open problems related​ to automated representation search, theoretical explanation of representation gain,​ cipher-family transferability, interpretability, reproducibility, and key-recovery integration.​ The survey highlights that rich input representations should be treated as first-class​ cryptanalytic design choices rather than secondary implementation details.

Note: Article Overview This paper presents a comprehensive survey and taxonomy of input representation strategies in neural differential cryptanalysis, a rapidly growing interdisciplinary field at the intersection of machine learning and symmetric-key cryptanalysis. Since the introduction of neural differential distinguishers for round-reduced SPECK, the research community has increasingly recognized that how cryptanalytic evidence is encoded and presented to neural networks substantially affects distinguishing performance, data efficiency, and practical attack relevance. Contribution and Significance Our work makes several novel contributions to the field: • Representation-Centric Framework: We introduce a unified framework that characterizes input representations through five components: difference set (D), number of observations per sample (m), sharing structure (s), encoding function (ϕ), and ciphertext cost (c). This framework provides common terminology for comparing diverse representation approaches. • Comprehensive Taxonomy: We organize existing neural distinguisher inputs into six distinct families: single-pair, multi-pair, multi-difference/matrix-style, multi-round/multi-splicing, structured encoding, and score-aggregation based representations. This taxonomy clarifies the evolution from simple to complex input designs. • Comparative Literature Analysis: We map representative works according to representation structure, target cipher family, design objectives, and attack relevance, identifying trends and gaps in the current literature. • Cost-Aware Evaluation Methodology: We analyze the critical distinction between fixed-sample and fixed-ciphertext comparisons, arguing that representation-rich distinguishers require explicit cost models to ensure fair evaluation and meaningful cryptanalytic interpretation. • Future Research Agenda: We identify and systematically discuss open problems, including automated representation search, theoretical foundations of representation gain, cipher-family transferability, interpretability of learned evidence, standard benchmarks, and integration into key-recovery attacks.