Secure Quantum-Based Adder Design for Protecting Machine Learning Systems Against Side-Channel Attacks

Abstract

Machine learning (ML) has recently been adopted in various application domains. Usually, a well-performing ML model relies on a large volume of training data and powerful computational resources. Recently, hardware accelerators utilizing field programmable gate arrays (FPGAs) have been developed to provide high-performance hardware while maintaining the required accuracy for ML tools. However, one of the main challenges hindering the FPGA-based ML models is their susceptibility to adversarial attacks, such as physical side-channel attacks. In this study, various kinds of countermeasures, including masking and hiding techniques, are examined to mitigate the aforementioned shortcomings and enhance the security of FPGA-based ML systems. In addition to FPGA-based defenses, the advantages of quantum computing for designing circuits to enhance data protection are also elaborated. However, concerning FPGA-based ML models, which are used to defend against physical side-channel attacks, quantum dot cellular automata (QCA) offers a more promising option. Its inherent security, lower power consumption, higher speed, and reduced vulnerability to side-channel leakage make it the best alternative. Therefore, this study emphasizes the implementation of the quantum nature of QCA to protect valuable information against physical side-channel attacks. It also offers quantum masking circuits for protecting sensitive information in machine learning systems, including XOR, adder, and RCA. Furthermore, the presented work advocates for leveraging QCA technology to augment the security of machine learning systems by mitigating the disclosure of sensitive data. The proposed QCA-based masked designs, which include an adder and a ripple carry adder (RCA), pose some qualities, which include a single-layer structure, minimal cell count, and low latency. When compared with the best counterparts among the recommended designs, these designs exhibit significant improvements regarding cell consumption and occupied area, with improvements of 33.3% and 36.6% respectively.