论文标题
SCA的聚类与统计分析：机器学习更好
Clustering versus Statistical Analysis for SCA: when Machine Learning is Better
论文作者
论文摘要
评估对SCA攻击的实施加密算法的阻力，以及在设计过程的早期阶段检测SCA泄漏源，对于有效的实施重新设计非常重要。因此，不依赖加密操作中处理的关键的有效SCA方法是有益的，并且可以成为实施加密方法的有效设计方法的一部分。在这项工作中，我们比较了两种用于分析椭圆曲线点乘法的功率痕迹的不同方法。与平均值进行比较的第一种方法是基于统计分析的简单方法。第二个是K-均值 - 用于数据群集的最多使用的无监督的机器学习算法。我们早期工作的结果表明，机器学习算法并不优于简单方法。在这项工作中，我们专注于使用两种分析方法的攻击结果进行比较，以了解其收益和缺点。我们的结果表明，与平均值的比较仅当在攻击的KP执行过程中处理的标量是平衡的，即标量k中的“ 1”的数量大约高于“ 0”的数量。与此相比，如果标量高度不平衡，K-均值也有效。即使标量k仅包含非常少数的“ 0”位，它仍然有效。
Evaluation of the resistance of implemented cryptographic algorithms against SCA attacks, as well as detecting of SCA leakage sources at an early stage of the design process, is important for an efficient re-design of the implementation. Thus, effective SCA methods that do not depend on the key processed in the cryptographic operations are beneficially and can be a part of the efficient design methodology for implementing cryptographic approaches. In this work we compare two different methods that are used to analyse power traces of elliptic curve point multiplications. The first method the comparison to the mean is a simple method based on statistical analysis. The second one is K-means - the mostly used unsupervised machine learning algorithm for data clustering. The results of our early work showed that the machine learning algorithm was not superior to the simple approach. In this work we concentrate on the comparison of the attack results using both analysis methods with the goal to understand their benefits and drawbacks. Our results show that the comparison to the mean works properly only if the scalar processed during the attacked kP execution is balanced, i.e. if the number of '1' in the scalar k is about as high as the number of '0'. In contrast to this, K-means is effective also if the scalar is highly unbalanced. It is still effective even if the scalar k contains only a very small number of '0' bits.
