Convolutional neural network attack on cryptographic circuits

In this Letter, a novel convolutional neural network (CNN) attack is explored as a new kind of malicious attack to disclose the secret key of cryptographic circuits. A cryptographic circuit with the known secret key is set as the reference model for training the CNNs. Moreover, a matrix that includes the information of an input plaintext and the known secret key of the cryptographic circuit is built as the input training data of the CNNs. A Sigmoid function and a step function are used to normalise and classify the power dissipation of the cryptographic circuit to generate the output training data of the CNNs, respectively. After training the CNNs, a cryptographic circuit with the unknown secret key can be cracked by hypothesising all the possible keys to test the well‐trained CNNs, because the correct secret key enables the CNNs to achieve the highest testing accuracy among all the hypothesised keys. As demonstrated in the results, the proposed CNN attack successfully reveals the secret key of a unprotected (protected) cryptographic circuit after analysing about 500 (100,000) data.