Machine Learning-based Detection of Attacks on AES Cryptography using FPGA Power Traces

The proposed work is aimed at the analysis of power traces of AES implementations on FPGA boards, using the application of machine learning techniques in the detection of side-channel attacks. We mainly concentrated on using the AES PTv2 dataset, for power traces collected from the Piñata board, under masked and unmasked AES encryption scenarios. In the proposed framework, we also considered data preprocessing, dimensionality reduction, and advanced machine learning models. Standardization and Principal Component Analysis effectively reduced the dataset to 300 components, with all important features being retained. To handle class imbalance, SMOTE was used in order to train the network, which is a way of balancing out the training with overfitting.The models that were considered include Random Forest, Multi-Layer Perceptron, and Convolutional Neural Network. Random Forest classifier reached 89% accuracy, along with good precision-recall balance. The optimized MLP model attained 81% accuracy with three hidden layers of neurons 256, 128, and 64. However, the CNN-based model outshone both these models by recording an impressive 99% accuracy with perfect precision, recall, and F1-scores. The results conclude that deep learning is very successful in extracting relevant spatial and temporal features for analysis in side channels and that it is better used in CNN, which can actually detect AES attacks. Additionally, we analyze the computational cost of each model and discuss practical limitations such as dependency on hardware variations and noise interference in power traces.