Scholarlink

This chapter presents attacks targeting unprotected devices. Specifically, devices may in fact be protected, but attacks will not be optimized based on the knowledge of the instantiated protection (such as masking). We detail both unprofiled and profiled attacks. While profiled attacks require a preliminary training before they can be carried out, unprofiled attacks can be carried out from scratch: they require only a single strategy (no learning strategy required, just guessing) and a set of leakage measures on which they can be applied. The strategy itself is the subject of this chapter. While profiled attacks require an open clone device that can be provisioned with a chosen key, one of the main attractions of unprofiled attacks is that they can work right out of the box. This presents an advantage in terms of practical applicability. Naturally, we begin this chapter with unprofiled attacks. The general side-channel analysis case is tackled with in Sect. 3.1, where we derive the maximum likelihood distinguisher. Then, we present correlation and mutual information unprofiled attacks in Sect. 3.2 and emphasize in which situation they are the most suitable. Interestingly, each one of them is proved “optimal” in two different, albeit real and representative, contexts. Afterward, we study in Sect. 3.3 optimal collision attacks, where the model is taken from the traces themselves, leveraging the fact that the same computation is executed (albeit under different data) multiple times. We continue in Sect. 3.4 with leakage agnostic metrics, namely, Kolmogorov-Smirnov (KS) distinguishers, which are independent of the exact leakage model (the KS analysis depends only on leakage ordering, not on values). Adaptation of the former case to the situation where the data dimensionality is large is the topic of Sect. 3.5. Similarly, when the masking order is too large, realistic distinguishers from a computational standpoint are explored in Sect. 3.6. Finally, Sect. 3.7 explains how side-channel leakage can be analyzed to reconstruct an explainable model.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Attacks on Unprotected Devices

  • 章节

  • 发表日期：

    2025年

查看全文
  • Wei Cheng,
  • Sylvain Guilley,
  • Olivier Rioul

摘要

This chapter presents attacks targeting unprotected devices. Specifically, devices may in fact be protected, but attacks will not be optimized based on the knowledge of the instantiated protection (such as masking). We detail both unprofiled and profiled attacks. While profiled attacks require a preliminary training before they can be carried out, unprofiled attacks can be carried out from scratch: they require only a single strategy (no learning strategy required, just guessing) and a set of leakage measures on which they can be applied. The strategy itself is the subject of this chapter. While profiled attacks require an open clone device that can be provisioned with a chosen key, one of the main attractions of unprofiled attacks is that they can work right out of the box. This presents an advantage in terms of practical applicability. Naturally, we begin this chapter with unprofiled attacks. The general side-channel analysis case is tackled with in Sect. 3.1, where we derive the maximum likelihood distinguisher. Then, we present correlation and mutual information unprofiled attacks in Sect. 3.2 and emphasize in which situation they are the most suitable. Interestingly, each one of them is proved “optimal” in two different, albeit real and representative, contexts. Afterward, we study in Sect. 3.3 optimal collision attacks, where the model is taken from the traces themselves, leveraging the fact that the same computation is executed (albeit under different data) multiple times. We continue in Sect. 3.4 with leakage agnostic metrics, namely, Kolmogorov-Smirnov (KS) distinguishers, which are independent of the exact leakage model (the KS analysis depends only on leakage ordering, not on values). Adaptation of the former case to the situation where the data dimensionality is large is the topic of Sect. 3.5. Similarly, when the masking order is too large, realistic distinguishers from a computational standpoint are explored in Sect. 3.6. Finally, Sect. 3.7 explains how side-channel leakage can be analyzed to reconstruct an explainable model.