區域性熱能洩漏之旁通道分析的改進

Abstract

本研究使用機器學習於旁通道攻擊，並提出一種應用於深度學習模型的訓練方法，迭代遷移式學習(iterative transfer learning)，目前大多數研究的旁通道攻擊，在進行破解 advanced encryption standard-128 (AES-128) 時，針對每一個位元組都是重新訓練同樣的模型，這樣的訓練方式未考慮位元組的相關性，但由於每一個位元組的訓練模式是非常的相似，因此本研究提出迭代遷移式學習，先訓練其中一個位元組，再以此模型作為預訓練模型 (pretrained model)，接續訓練其餘位元組，這樣的方式能夠在訓練其餘位元組時，提供好的初始權重，並減少模型攻擊階段的 measurement-to-disclosure (MTD)。在訓練資料充足的情況下，使用迭代遷移式學習能夠些微的減少 MTD，搭配多層感知器能夠將平均 MTD 從 54.7 減少至 53.8，搭配卷積神經網路能夠將平均 MTD 從 125 減少至 82.7，在訓練資料不足的情況下，使用迭代遷移式學習依然能夠成功破解加密，根據實驗結果顯示，此方法在減少訓練資料從13,600減少到2,000筆資料，仍然可以成功破解AES-128，迭代遷移式學習搭配多層感知器平均 MTD 為 635，而常見分開訓練的方式則是無法成功的。除此之外，本研究也比較使用功耗資料和溫度變化資料，進行訓練和攻擊的差異，結果顯示，使用功耗資料訓練的模型的平均 MTD 大部分優於使用溫度訓練的模型。

The research uses machine learning for side-channel attacks and proposes a training approach applied to deep learning models called iterative transfer learning. Currently, most of the side-channel attacks train a model for each bytes, but this training method does not consider the correlation of each bytes. Because the training of each byte is very similar, the research proposes iterative transfer learning. In the beginning, using a byte to train the model and use this model as a pretrained model to train the remaining bytes. This approach can provide good initial weights when training the remaining bytes and reduce the measurement-to-disclosure (MTD) in the attack phase. When the data is enough, using iterative transfer learning still can reduce the MTD slightly. Using iterative transfer learning with multilayer perceptron can reduce average MTD from 54.7 to 53.8. Using iterative transfer learning with convolution neural network can reduce average MTD from 125 to 82.7. When the data is insufficient, using iterative transfer learning still can decrypt the MTD successfully. According to the experiments, this approach can successfully crack AES-128 when the number of training data is reduced from 13,600 to 2,000, while the separate training approach cannot be successful. And the average MTD is 635 when using iterative transfer learning with multilayer perceptron. Besides, we also examine the difference between using the power data and thermal data to attack the AES-128. In our research, we find that the average MTD of the model trained with power data is mostly better than the model trained with thermal data.