Role of Machine Learning in Power Analysis Based Side Channel Attacks on FPGA

The cloud-based devices face many threats these days due to shared resources like power. The attacker measures the power by using some remote sensors which are present in attacker tenants. These sensors get partial or full access to a power distribution network (PDN) and work as a backdoor for the attacker. In our research, we explored potential security issues which involved power analysis-based side-channel attacks (SCAs) on Field Programmable Gate Arrays (FPGAs). We have made three major contributions to our research paper. First, we have discussed the power analysis or power profiling of FPGA, which is dependent upon voltage fluctuations' leakage while performing some encryption tasks. The voltage fluctuations of the cryptographic module are measured by some physical source like an oscilloscope or remote source like delay line sensors. Second, we have discussed potential power analysis-based SCAs that used these measurements of voltage fluctuations to extract the secret key. Third, we have designed a framework based on machine learning (ML) and deep learning (DL) models to perform secret key predictions and successful attacks. Firstly, our custom convolutional neural networks (CNN) model has revealed all 16 bytes of the secret key and performed a successful attack with only 570 attack power traces. Secondly, the multi-layer perceptron (MLP) model has successfully attacked only using 3200 traces using the same framework. Overall we have achieved a better performance in terms of the required number of power traces for a successful attack, training time, prediction time, and attack time.