Beyond encryption: How deep learning can break microcontroller security through power analysis

Abstract

• Development of a novel deep learning-based power analysis attack for extracting AES encryption algorithm keys from the ATmega328P microcontroller. • Utilization of a convolutional neural network (CNN) implemented with Python/Keras, trained on power consumption traces collected during AES-128 execution on the ATmega328P microcontroller. • Creation of a CNN-based Power Analysis Attack (PAA) specifically designed for recovering AES keys from microcontrollers, using a large dataset of 150,000 power consumption traces. • Practical insights gained from implementing the CNN-based PAA on a 130 nm ASIC, highlighting the application of machine learning in cybersecurity. • Successful key extraction with minimal traces (around 1100), indicating a significant improvement over state-of-the-art techniques. This paper investigates the application of convolutional neural networks (CNNs) for power analysis attacks (PAAs) on cryptographic systems, specifically targeting resource-constrained devices like microcontrollers. Vulnerabilities in these systems stem from unintended information leakage through side channels, such as power consumption during cryptographic operations. By utilizing CNNs, attackers can analyze these measurements to potentially extract secret keys. We propose a CNN-based PAA designed to recover Advanced Encryption Standard (AES) keys from microcontrollers. The CNN was trained on a dataset of 150,000 power consumption traces collected during AES encryption. This paper explores how our CNN-based method exploits information leakage to recover secret keys and compares its performance against existing approaches. Our method, implemented on an ASIC with 130 nm technology, successfully extracts keys using just 1100 traces, marking a substantial improvement over current state-of-the-art technique.