Improved Hybrid Approach for Side-Channel Analysis Using Efficient Convolutional Neural Network and Dimensionality Reduction

Abstract

Deep learning-based side channel attacks are burgeoning due to their better efficiency and performance, suppressing the traditional side-channel analysis. To launch the successful attack on a particular public key cryptographic (PKC) algorithm, a large number of samples per trace might need to be acquired to capture all the minor useful details from the leakage information, which increases the number of features per instance. The decreased instance-feature ratio increases the computational complexity of the deep learning-based attacks, limiting the attack efficiency. Moreover, data class imbalance can be a hindrance in accurate model training, leading to an accuracy paradox. We propose an efficient Convolutional Neural Network (CNN) based approach in which the dimensionality of the large leakage dataset is reduced, and then the data is processed using the proposed CNN based model. In the proposed model, the optimal number of convolutional blocks is used to build powerful features extractors within the cost limit. We have also analyzed and presented the impact of using the Synthetic Minority Over-sampling Technique (SMOTE) on the proposed model performance. We propose that a data-balancing step should be mandatory for analysis in the side channel attack scenario. We have also provided a performance-based comparative analysis between proposed and existing deep learning models for unprotected and protected Elliptic curve (ECC) Montgomery Power ladder implementations. The reduced network complexity, together with an improved attack efficiency, promote the proposed approach to be effectively used for side-channel attacks.

Machine Learning Attacks and Countermeasures on Hardware Binary Edwards Curve Scalar Multipliers

“…A straightforward counter against these type of SCAs is the utilization of SM algorithms that are highly regular, i.e., the scalar bit that is being processed in each round does not appear to have a link to the round's operations (e.g., MPL or Double and Always Add) [16]. Nevertheless, there exist some SCAs that have the ability to thwart this regularity, mainly by altering the EC base point used as an input of the SM [20]. For this objective, Comparative SCAs were introduced (initially based on Power Attacks but adapted to Electromagnetic Emission attacks as well).…”

“…The goal of masking is the decoupling of sensitive data from their corresponding leakage trace. Crucial for this method is the existence of some randomization technique (multiplicative or additive) upon the secret information that has a sensitivity to leakage O i (t) [20]. Three well-established EC SM masking techniques have been used in practice (appearing in several variations), originally proposed by Coron in [33].…”

“…With ML, an adversary does not need to create a perfect leakage model (using all possible inputs of a secret value block) but instead allows an ML model to be built with a non-exhaustive series of leakage traces (that can only be corresponded to specific, instead of all, secret block values). As the leakage becomes more noisy (either due to the hiding or the masking countermeasures implemented), more accurate results can be extracted by applying an approach based on ML profiling [9,10,20].…”

“…In [20], a time-efficient Convolutional Neural Network (CNN)-based approach is realized, focusing on dimensionality reduction in order to handle the high computational complexity of CNN attacks. In the proposed CNN model, the optimal number of convolutional blocks is used to build powerful feature extractors within the cost limit.…”

“…However, ref. [20] does not assess the ECC SM implementations against small datasets and simple Machine Learning algorithms. Thus, in order to form a complete performance analysis for small and medium sized datasets, this paper uses two supervised Machine Learning algorithms and one simple artificial neural network, complementing and extending the findings in [20].…”