Deep learning-based profiling side-channel attacks in SPECK cipher.

Over recent years, deep learning-based profiling side-channel analysis has garnered significant attention in both academia and industry. Research in this field has progressed to address challenges such as hyperparameter tuning in deep learning models, with ensemble methods emerging as a noteworthy approach. However, these techniques have primarily been applied to classical symmetric ciphers like AES, while their adaptation to lightweight cryptographic primitives remains in its infancy and has been rarely explored. SPECK, a lightweight cipher developed by the National Security Agency, is designed specifically for resource-constrained devices. Its widespread adoption in the Internet of Things (IoT) highlights the critical need for robust protection against side-channel attacks. This paper introduces a novel profiling side-channel analysis technique leveraging deep learning for the SPECK cipher. We demonstrate the effectiveness of our approach by employing sequential divide and conquer ensemble of deep learning models to attack software implementations of SPECK-32/64 cipher, successfully recovering its 8-byte secret key in fewer than 250 traces. To the best of our knowledge, this work represents the first deep learning-based profiling attack on an unprotected and protected implementations of SPECK.