Using Deep Learning Attribution Methods for Fault Injection Attacks

Explore a groundbreaking 40-minute conference talk that demonstrates the first-ever application of deep learning attribution methods from image processing as reverse engineering tools for fault injection attacks on cryptographic implementations. Learn how to leverage gradient and layer-wise relevance propagation (LRP) techniques to analyze power consumption traces from embedded systems, specifically targeting a secure EEPROM (Analog Devices DeepCover DS28C36) using a black box approach. Discover the methodology for collecting power consumption data during protected and unprotected memory read operations, training deep learning models to identify differences between these states, and using attribution methods to reverse engineer the model's decision-making process. Understand how this analysis reveals critical timing information for manipulating security fuses and uncovers countermeasures such as double checking mechanisms designed to prevent single fault injection attacks. Master the practical implementation of double laser fault injection techniques that can bypass multiple security checks to extract protected EEPROM secrets, representing a significant advancement in hardware security research that extends deep learning applications beyond traditional side-channel attacks to fault injection scenarios.