Deep Learning-based Side-channel Analysis (DL-SCA) has emerged as the state-of-the-art technique, surpassing traditional SCA methods that are based on statistical analysis or template profiling by leveraging complex leakage signals for secret retrieval. In DL-SCA, an adversary trains a deep learning (DL) model using data from a profile device under their control and subsequently employs this trained model to recover the secret key from a target device. However, in realistic attack scenarios, the profile device may differ from the target device, and usually only unlabeled leakage traces of the target device are available due to limited control by the adversary, making supervised learning infeasible. This project aims to develop a realistic universal DL-SCA approach utilizing unsupervised transfer learning. Our novel methods enable transferring DL-SCA across various implementation platforms and across various side channels, including power, electromagnetic (EM), and cache-based channels.