Although side-channel attacks based on deep learning are widely used in AES encryption algorithms, there is little research on lightweight algorithms. Lightweight algorithms have fewer nonlinear operations, so it is more difficult to attack successfully. Taking SPECK, a typical lightweight encryption algorithm, as an example, directly selecting the initial key as the label can only crack the first 16-bit key. In this regard, we evaluate the leakage of SPECK's operations (modular addition, XOR, shift), and finally select the result of XOR operation as the label, and successfully recover the last 48-bit key. Usually, the divide and conquer method often used in side-channel attacks not only needs to train multiple models, but also the different bytes of the key are regarded as unrelated individuals. Through the visualization method, we found that different key bytes overlap in the position of the complete electromagnetic leakage signal. That is, when SPECK generates a round key, there is a connection between different bytes of the key. In this regard, we propose a transfer learning method for different byte keys. This method can take advantage of the similarity of key bytes, improve the performance starting-point of the model, and reduce the convergence time of the model by 50%.
Keywords: SPECK; deep learning; intermediate operation; side channel analysis; transfer learning.