I. Introduction
Side-channel Analysis (SCA) has been widely used to evaluate the security level of hardware devices using Physical Leakage Information (PLI) such as power dissipation or Electro-Magnetic (EM) emanation. Mostly, SCA targets the secret key used in block ciphers like Data Encryption Standard (DES), Advanced Encryption Standard (AES) implemented in Hardware. Recently, Deep learning (DL) starts to be adopted to SCA, but mostly in the field of profiling SCA [1], [2], [3], [4], [5], where a copy of the target device is available to be used for training. However, the application of DL in non-profiling SCA, where the copy of the target device is not available, is scarce. Among the reported non-profiling SCA based on DL, the state-of-the-art is the reported Correlation Optimization with Deep Learning Analysis (CO-DLA) [6], which has been proven to be able to suppress SCA countermeasures like ‘Masking’ [7], [8], [9], [10] and ‘Hiding’ [11], [12], [13], [14]. In CO-DLA, neural networks are used to compress each PLI measurement into one informative output, meaning that the output possesses strong correlation with the Hypothesis Leakage Model (HLM), such as Hamming Weight (HW) and Hamming Distance (HD). These HLMs are computed using the secret key and the plaintext or ciphertext associated with the compressed PLI measurement. Despite having a strong capability to extract the interdependency between the HLM and PLI measurements, CO-DLA requires training one DL model for each possible key guess, which is a time-consuming process. Furthermore, CO-DLA was not investigated with any preprocessing techniques before, as DL is believed to be able to cope with any complex input. In this paper, we propose a novel Wavelet Scattering Transform (WST) based CO-DLA to reveal secret keys with less time and fewer PLI measurements. With a properly chosen preprocessing technique to decompose complex input, the workload of the neural network in CO-DLA is reduced, hence reducing the training time and PLI measurements required for secret key extraction.