Boosting the efficacy of power attacks on cryptographic circuits with autoencoder

Power attacks are able to leak the secret key of protected cryptographic circuits as long as the number of sampled power data is sufficient to filter the injected noise. However, in this Letter, a novel power attack is proposed to break protected cryptographic circuits when the number of available power data is inadequate for filtering the injected noise. To achieve this goal, a state‐of‐the‐art machine learning technique: autoencoder is utilised for generating a certain number of new power data that are similar to the original power data to assist the noise filtering. Within the proposed autoencoder neural networks, the input and output layers are set with 1024 × 1 arrays. Furthermore, two convolution layers are used in the encoder block and decoder block, respectively, to achieve similar features between the input layer and the output layer. Under the assistance of autoencoder, the novel power attack reveals the secret key of a protected cryptographic circuit with 300,000 power data. In contrast, the regular power attacks fail to break the cryptographic circuit under the same number of power data.