Bias-variance Decomposition in Machine Learning-based Side-channel Analysis
Bias-variance Decomposition in Machine Learning-based Side-channel Analysis
复制标题
基于机器学习的侧信道分析中的偏差-方差分解
DOI：
--
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
S. Picek
中科院分区：
文献类型：
--
作者：
Daan van der Valk;S. Picek
机器学习技术是侧通道分析中的一个强大选项。尽管如此，在许多情况下，它们的性能远未达到预期。在这样的场合，理解问题的难度和机器学习算法的行为是非常重要的。为此，人们不仅需要研究机器学习的性能，还需要深入了解其可解释性。使我们能够做到这一点的一个工具是偏差-方差分解，我们能够将预测误差分解为偏差、方差和噪声。通过这种技术，我们可以分析各种场景，并识别问题的困难来源，以及额外的测量/特征或更复杂的机器学习模型如何缓解问题。虽然这样的结果是有希望的，但仍然存在缺点，因为通常不容易将侧信道攻击的性能和由偏差方差分解给出的机器学习分类器的性能联系起来。在本文中，我们提出了一种新的工具来分析基于机器学习的侧信道攻击的性能-猜测熵偏差方差分解。通过它，我们能够更好地了解各种机器学习技术的性能，并了解设置的变化如何影响攻击的性能。为了验证我们的说法，我们给出了大量的实验结果，为一些不同的设置。
Machine learning techniques represent a powerful option in profiling sidechannel analysis. Still, there are many settings where their performance is far from expected. In such occasions, it is very important to understand the difficulty of the problem and the behavior of the machine learning algorithm. To that end, one needs to investigate not only the performance of machine learning but also to provide insights into its explainability. One tool enabling us to do this is the bias–variance decomposition where we are able to decompose the predictive error into bias, variance, and noise. With this technique, we can analyze various scenarios and recognize what are the sources of problem difficulty and how additional measurements/features or more complex machine learning models can alleviate the problem. While such results are promising, there are still drawbacks since often it is not easy to connect the performance of side-channel attack and performance of a machine learning classifier as given by the bias-variance decomposition. In this paper, we propose a new tool for analyzing the performance of machine learning-based side-channel attacks – the Guessing Entropy Bias–Variance Decomposition. With it, we are able to better understand the performance of various machine learning techniques and understand how a change in a setting influences the performance of an attack. To validate our claims, we give extensive experimental results for a number of different settings.