Wireless communication networks require robust authentication meth-ods to safeguard against unauthorized access and potential threats.Traditional authentication techniques, such as cryptographic hand-shakes, while effective, have limitations, including increased networkoverhead and vulnerability of secret keys. RF fingerprinting presentsan innovative solution by utilizing the unique, hardware-induced im-perfections in radio frequency (RF) circuitry to create distinctive "fin-gerprints" for device authentication. This method reduces reliance onsecret keys, streamlines the authentication process, and enhances se-curity by enabling automatic device recognition based on their uniqueRF signal patterns.This thesis investigates the use of RF fingerprinting in wireless net-works, focusing on advanced deep-learning techniques for enhanc-ing identification and security. Conducted as part of the EU projectCORENext, which aims to develop sustainable and trustworthy Be-yond 5G (B5G) and Sixth Generation (6G) technologies, the researchleverages advanced deep learning models, including ConvolutionalNeural Networks (CNNs) and Long Short-Term Memory (LSTM) net-works, to classify and authenticate devices. The study introduces theMixture of Experts (MOE) and Multi-Task Learning (MTL) models.Additionally, we design a baseline model with a new novel customloss function to improve accuracy and robustness. This loss functioncombines Focal Loss, which adjusts the learning process by givingmore weight to examples that are harder to classify, thereby enhanc-ing the accuracy of device identification based on unique RF signa-tures. Additionally, it incorporates a dynamic penalty mechanism us-ing the confusion matrix to minimize false positives and false nega-tives. This adjusts penalties based on the model’s errors, giving higherpenalties to classes with more false positives and false negatives. Thecombined approach of Focal Loss and confusion matrix-based penal-ties enables the model to maintain consistent accuracy across variedRF conditions, as it continuously adapts to the complexity and dis-tribution of the data. By tailoring data segmentation parameters andevaluating model performance, our thesis aims to identify the mostsuitable approach for RF fingerprinting in diverse and dynamic wire-less environments, contributing to the advancement of secure wire-less communication technologies.