Active Side Channel Analysis for Cross-Device Attack

Tuan Dung Pham, Viet-Hung Tran, Zichi Zhang, Ngoc Phu Doan, Tuan Hoang, Hans Vandierendonck, Maire O'Neill, Son Thai Mai

Keywords: ML and AI security and privacy, ML and AI applications to security and privacy, Hardware security, Side channels

TL;DR: ActSCA: Active Side Channel Analysis for Cross-Device Attack

Abstract:

Side Channel Analysis (SCA) exploits relationships between physical signals of a device and its actual computation to extract sensitive information, causing serious threat to privacy and security. Among various approaches, Deep Learning-based profiling attacks (DL-SCA) have recently emerged as one of the most powerful methods due to their ability to fully characterize the target devices. However, they suffer from major drawbacks including huge data consumption and lack of portability across different target devices. This paper introduces Active SCA (ActSCA), a $unique$ and $generic$ framework for boosting performance of any base DL-SCA model.
ActSCA fundamentally differs to existing research as follows.
Firstly, rather than relying on large training data in the profiling stage, it $actively$ selects subsets of training data and $iteratively$ refine the model to avoid overfitting, thus enhancing performance. Secondly, in the attack stage, ActSCA $exploits$ existing training data pool from profiling devices to construct $separate$ attack models for different target devices without requiring any training data from the attacking devices as is the case in other existing methods by using only few unlabeled SCA traces collected during the attacking phase to guide the model adaptation process.
These make ActSCA a highly $portable$ and $practical$ attack method. We demonstrate its performances on the Post Quantum (PQC) Kyber algorithm using power leakage to retrieve secret keys. ActSCA significantly improves the performances of all employed base models and outperforms all recent approaches like CNNC, MDMSD, ZMUV, MMD, ADA in terms of mean rank and top-$k$ accuracy.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 12464