SCARL: Side-Channel Analysis with Reinforcement Learning on the Ascon Authenticated Cipher
This addresses a key bottleneck in cryptographic security analysis by enabling side-channel attacks without prior leakage models, though it is incremental as it builds on existing reinforcement learning and auto-encoder techniques.
The paper tackles the problem of side-channel analysis requiring prior leakage models by introducing SCARL, an unsupervised reinforcement learning method that extracts data-dependent features from power measurements, recovering the secret key of the Ascon cipher using 24K traces, while classical methods fail with over 40K traces.
Existing side-channel analysis techniques require a leakage model, in the form of a prior knowledge or a set of training data, to establish a relationship between the secret data and the measurements. We introduce side-channel analysis with reinforcement learning (SCARL) capable of extracting data-dependent features of the measurements in an unsupervised learning approach without requiring a prior knowledge on the leakage model. SCARL consists of an auto-encoder to encode the information of power measurements into an internal representation, and a reinforcement learning algorithm to extract information about the secret data. We employ a reinforcement learning algorithm with actor-critic networks, to identify the proper leakage model that results in maximum inter-cluster separation of the auto-encoder representation. SCARL assumes that the lower order components of a generic non-linear leakage model have larger contribution to the leakage of sensitive data. On a lightweight implementation of the Ascon authenticated cipher on the Artix-7 FPGA, SCARL is able to recover the secret key using 24K power traces during the key insertion, or Initialization Stage, of the cipher. We also demonstrate that classical techniques such as DPA and CPA fail to identify the correct key using traditional linear leakage models and more than 40K power traces.