Elie Bursztein - Deep Learning: Revolutionizing Side Channel Cryptanalysis
Demonstrates how deep learning models outperform classical DPA attacks for power-analysis-based key extraction from AES implementations.
Summary
Deep Learning for Side Channel Attacks
Elie Bursztein (Lead Security Researcher, Google) presents research applying deep neural networks to power side-channel analysis, achieving superior key extraction performance compared to classical Differential Power Analysis (DPA).
Classical vs. ML-Based SCA
Classical DPA requires profiling attacks with a known device or relies on univariate statistical models. Neural network models trained on power traces directly learn the mapping from trace shape to intermediate computation values, handling misalignment and noise that breaks classical methods.
Convolutional Neural Network Architecture
Bursztein's approach uses CNNs to process raw power traces, with the model learning alignment-invariant features automatically. On STM32 AES implementations, the CNN recovers full 128-bit AES keys with fewer traces than classical CPA.
Open-Source Dataset and Framework
The talk releases a public power trace dataset and open-source framework for training and evaluating SCA models, enabling reproducible research.
Related talks
Main Stage·44 min
Ben Nassi · Ofek Vayner
Recovers cryptographic keys by analyzing video footage of device power LEDs, using camera rolling shutters to measure CPU power consumption fluctuations.
Main Stage·43 min
Elie Bursztein
Google researcher Elie Bursztein combines deep learning side-channel attacks with AI explainability to pinpoint and remediate cryptographic hardware leakage.
Main Stage·33 min
Elie Bursztein
Elie Bursztein discusses deep learning side-channel analysis, AI explainability tools, and hardware leak testing in a live Q&A.