<p>Deep learning techniques provide the possibility to skip traditional feature engineering in side-channel analysis. Recent papers have shown that deep learning can handle protected implementations without preprocessing and allow for implicit feature selection. However, feature engineering still represents the most important stage of side-channel analysis. Considering the extremely lengthy traces, it is evident that performing complete attacks poses significant challenges. To this end, we investigate feature selection methods from the machine learning domain to improve the performance of profiling attacks. A novel mutual information feature selection method is proposed, called FSML. Our method adopts label distribution learning to extract high-quality features, which improve leakage assessment under the worst-case security scenario, resulting in powerful attacks. We validate our method on two public datasets, ASCADr and ASCADv2. The results demonstrate that using small neural networks can outperform the current state-of-the-art results. Moreover, our method enhances the effectiveness of second-order profiling attacks and performs well even on high-noise datasets, proving its robustness against noise interference.</p>