Machine Learning-based Side-Channel Analysis of Cryptographic Chips
Advances in communication and network systems have given rise to a newera of revolutionary technology known as Internet of Things (IoT): a world where everything is connected through uniquely identifiable intercommunicating smart devices using uninterrupted connectivity and sensors. IoT brings convenience to human lives in almost all sectors, including healthcare centers, smart homes, and traffic management. The smart devices deployed in IoT systems consist of resource constraint embedded chips for processing private information. The security of the information processing on these chips has increased with the introduction of Edge Artificial Intelligence where information will be processed locally on the edge devices instead of the cloud.
Generally, on cryptographic chips, private information is secured using the standard mathematically strong cryptographic algorithms but the weak implementations of these algorithms can lead to side-channel leakages, which can be exploited to retrieve the secret information, hence endangering the user’s privacy and data. Moreover, the efficiency of these attacks has greatly improved due to the introduction of machine learning techniques which has even weaken the countermeasure-protected implementations. Despite being effective, such machine learning-based attacks have their own challenges like over-fitting due to redundant features , requirement of huge datasets for learning the leakage patterns, or being prone to produce inaccurate analysis because of low instance-feature ratio or imbalance classes. Moreover, these attacks are generally analyzed for symmetric ciphers, while asymmetric ciphers remain unaddressed.
The aim of this dissertation is to propose the improved machine learning-based sidechannel attacks using evolving machine learning technologies which can aid in recovering the sensitive information efficiently. To achieve this, we have designed and developed two novel hand-crafted feature engineering techniques to eliminate the redundant features, hybrid deep learning-based scheme for data with low instance-feature ratio, neural network-based model for various side-channel attack models, and a Generative Adversarial Network-based model to increase the dataset size by generating fake leakages. Moreover, a generalized few-shot learning-based leakage assessment model is proposed which combines the leakages from the multiple sources and channels to detect and differentiate the secret information in leakages. Furthermore hyperparameter tuning is performed to select the best models.
The comparison of the proposed attack models/schemes is performed with the state-ofthe- art side-channel attacks and with the other machine learning techniques. To evaluate the efficiency of the developed models, experiments are conducted on the protected and unprotected algorithms’ implementations leakages of both symmetric (AES) and asymmetric (ECC) ciphers on various cryptographic chips. The results demonstrate that the proposed methods outperform the state-of-the-art side-channel template attacks and can recover the sensitive information with better efficiency. It is concluded that machine learning-based sidechannel attacks pose a significant threat to the security of the cryptographic chips. Based on the findings, we suggest that new countermeasures should be designed which are effective against these advance attacks to secure user information.