Further observations on the security of Speck32-like ciphers using machine learning

Area of Science:

• Cryptography
• Machine Learning
• Cybersecurity

Background:

• Internet of Things (IoT) deployment necessitates robust security for constrained devices.
• Lightweight cryptography offers solutions for resource-limited environments.
• Evaluating cryptographic primitives, like Speck32, is crucial for secure communications.

Purpose of the Study:

• To develop a machine learning-driven framework for evaluating rotation parameters in Speck32-like ciphers.
• To assess the security implications of different rotational parameter choices.
• To identify optimal parameters for enhanced resistance against cryptanalysis.

Main Methods:

• Established a machine learning security evaluation framework for rotational parameters.
• Developed neural-differential distinguishers utilizing low-Hamming-weight and optimal differential input models.
• Evaluated 256 rotation parameters based on neural distinguisher accuracy.

Main Results:

• The (7,3) rotation parameter exhibited greater resilience against machine learning-aided distinguishing attacks than the standard (7,2) parameter.
• Identified bit bias in output and truncated differences as key factors influencing distinguisher accuracy.
• This work represents the first comprehensive ML-based security assessment of Speck32-like cipher rotation parameters.

Conclusions:

• The (7,3) parameter offers improved security for Speck32-like ciphers against ML-based attacks.
• Understanding bit bias is essential for designing more secure lightweight ciphers.
• The developed framework provides a novel approach for cryptographic parameter security evaluation.