Breaking the Blindfold - Deep Learning-based Blind Side-channel Analysis

Learn about a groundbreaking deep learning framework for conducting side-channel attacks without access to known plaintext or ciphertext in this 20-minute conference presentation from USENIX Security '25. Discover how the Deep Learning-based Blind Side-channel Analysis (DL-BSCA) framework overcomes the traditional limitations of physical side-channel analysis by leveraging deep neural networks to recover secret keys in blind scenarios. Explore the innovative Multi-point Cluster-based (MC) labeling method that accounts for dependencies between leakage variables by exploiting multiple sample points, significantly improving trace labeling accuracy. Examine validation results across four comprehensive datasets covering symmetric key algorithms including AES and Ascon, as well as the post-quantum cryptography algorithm Kyber, tested on platforms ranging from high-leakage 8-bit AVR XMEGA to noisy 32-bit ARM STM32F4 systems. Understand how this approach achieves the first successful blind side-channel analysis on desynchronization countermeasures, demonstrating practical effectiveness where previous methods failed to recover keys. Gain insights into real-world applications and the implications for cryptographic security assessment in scenarios where traditional side-channel analysis assumptions are invalidated, such as in Offset CodeBook encryption modes and complex hardware implementations.