Tue
17
Feb
Seminars and Conferences
Automatization of Differential Cryptanalysis
On Tuesday, 17 February 2026, at 2:30 pm, will take place Automatization of Differential Cryptanalysis, a seminar by Rocco Brunelli, from the “CrypTO Seminars” series, organised by the Department of Mathematical Sciences "Giuseppe Luigi Lagrange"-DISMA in collaboration with Telsy SpA, the TIM Group's center of expertise in cryptography and cybersecurity, operating within the scope of TIM Enterprise.
Abstract
The increasing complexity of modern block ciphers makes the manual design and analysis of cryptanalytic attacks progressively more difficult, motivating the use of automated and data-driven techniques. In recent years, machine learning has emerged as a promising tool for cryptanalysis, in particular for the construction of distinguishers on reduced-round primitives. In the first part of this seminar, I give an overview of how machine learning approaches cryptanalysis, focusing on neural distinguishers and their relationship with classical differential attacks. I discuss common design choices, threat models, and limitations of purely black-box learning-based methods. The second part of the seminar concentrates on a concrete machine-learning-based attack. I present a generic feature-engineering technique based on partial decryption, which incorporates structural information about the cipher into neural models. This approach improves the efficiency and robustness of neural distinguishers, while also enhancing their interpretability by linking learned features to well-known cryptanalytic properties. Overall, the seminar shows how machine learning can be leveraged as an effective and principled tool for cryptanalysis, bridging data-driven techniques and classical cryptographic insight.
The seminar is open to everyone and will take place in person.
Abstract
The increasing complexity of modern block ciphers makes the manual design and analysis of cryptanalytic attacks progressively more difficult, motivating the use of automated and data-driven techniques. In recent years, machine learning has emerged as a promising tool for cryptanalysis, in particular for the construction of distinguishers on reduced-round primitives. In the first part of this seminar, I give an overview of how machine learning approaches cryptanalysis, focusing on neural distinguishers and their relationship with classical differential attacks. I discuss common design choices, threat models, and limitations of purely black-box learning-based methods. The second part of the seminar concentrates on a concrete machine-learning-based attack. I present a generic feature-engineering technique based on partial decryption, which incorporates structural information about the cipher into neural models. This approach improves the efficiency and robustness of neural distinguishers, while also enhancing their interpretability by linking learned features to well-known cryptanalytic properties. Overall, the seminar shows how machine learning can be leveraged as an effective and principled tool for cryptanalysis, bridging data-driven techniques and classical cryptographic insight.
The seminar is open to everyone and will take place in person.