Research on AES Cryptographic Chip Electromagnetic Attack Based on Deep Transfer Learning

Abstract

Cryptographic chip is an important carrier for implementing cryptographic algorithms. Its side channel leakage seriously threatens information security, which has attracted extensive attention from researchers and business circles around the world. In this paper, the AES-128 cryptographic algorithm is used as the attack object implemented by the commonly FPGA cipher chip. The electromagnetic radiation detection is the main test method while the deep learning method is applied to the side channel attack. For traditional template attack method, its shortcoming is that the plaintext or ciphertext must be known. Combined with the “divide and conquer” approach, we proposed a nibble attack method that targets the key directly. This method can not only reduce the number of templates that need to be built, but also reduce the computational complexity. At the same time, this method can use more training data. For the problem of time-consuming training and large samples collection, transfer learning is used to assign initial weights to the training model. In this way, we fixed the AES-128 plaintext and keys and collected its electromagnetic data, in order to obtain an ideal template to transfer the deep learning situation to the specific case which the keys and plaintext are random to improve deep learning speed and accuracy of key attacks. The experimental results show that the recovery rate of the attack model to the nibble key can reach 77.13% while it has increased 9.47%. compared with the direct training. Besides, the training process is reduced by about 40 rounds, and the speed is increased by nearly 40%.