A Novel Attention-Based LSTM Model for Non-Profiled Side-Channel Attacks

Non-profiled side-channel attacks represent a powerful class of attacks for no excessive information from the target device is required. Recently, the application of deep learning techniques on non-profiled side-channel attacks has yielded promising results. However, some existing studies have not fully exploited the properties of power traces as a time series. In this paper, an attention-based Long Short-Term Memory (LSTM) model for non-profiled attacks is proposed. This novel architecture combines LSTM and an attention mechanism in order to precisely extract the features of time samples in long-term traces. Through experiments on a public dataset, this architecture outperforms previous methods by producing more distinguishable results in identifying the correct key. Moreover, this architecture demonstrates reliable results in attacking power traces with additional Gaussian noise.