Lightweight Cryptographic Models for IoT Devices: A Deep Learning Approach to Power Side-Channel Attack Prevention

The implementation of lightweight cryptography is often found in unrolled architecture, which offers the advantages of low latency and high real-time performance but also runs the risk of Side-Channel Attack (SCA). These days, the Internet of Things has led to a variety of applications that need lightweight cryptographic primitives, including block cyphers, for safe and effective computation with little resources. The expense of developing machine learning (ML) models makes them potentially trade secrets. They must thus be shielded against harmful types of reverse engineering (such as in IP piracy). As machine learning continues to move to edge devices, partly for performance reasons and partly for privacy reasons, the models are now vulnerable to what are known as physical side-channel assaults. Earlier studies have shown that power-based side-channel assaults may recover such control flow in highly restricted contexts, but they depended on significant changes in computational stages or data dependencies to differentiate between states in a state machine. Using Field Programming Gate Arrays (FPGAs), we investigated possible security vulnerabilities involving side-channel assaults (SCAs) based on power analysis. We have significantly improved our study report in three ways. The power analysis or power profile of FPGA, which depends on the leakage of voltage fluctuations during certain encryption activities, was covered first. A physical source, such as an oscilloscope, or a remote source, such as delay line sensors, are used to detect the fluctuations in voltage of the cryptography module. Second, we spoke about possible power analysis-based SCAs that extracted the secret key using these voltage fluctuation readings. Third, we have created a framework for successful assaults and secret key predictions that is based on machine learning (ML) and deep learning (DL) algorithms. First off, using only 570 attack power traces, our proprietary convolutional neural networks (CNN) model successfully executed an attack and exposed all 16 bytes of the secret key. Second, the same architecture has been used to effectively attack the multi-layer perceptron (MLP) model using only 3200 traces. In terms of training time, prediction time, attack time, and the amount of power traces needed for a successful attack, we have improved overall.