The assessment of a cryptographic implementation's worst-case security is critically reliant on the attacks that have been profiled. A substantial amount of attention has been paid to the development of profiled attacks during the last sixteen years. These attacks have included both deep learning-based and template-based attacks. On the other hand, information in the frequency domain can be lost since the bulk of attacks happen in the time domain. In this study, we propose a new side-channel attack in time–frequency representations based on deep learning to make greater use of leakage information. These representations are the target of the attack. In order to fully use convolutional neural networks in profiled attacks, we concurrently extract high-level key-related information from spectrograms and employ time–frequency patterns. This enables us to use these networks’ potential to the fullest. In order to execute successful attacks, it is first necessary to set up a functional network architecture. Second, several crucial spectrogram parameters are reviewed in order to improve the network's training. Additionally, utilizing available datasets in the temporal and time–frequency domains, we contrast CNN-based attacks with template attacks. The heuristic findings from these experiments provide a fresh viewpoint and show that CNN-based attacks in spectrograms could effectively replace state-of-the-art profiled attacks. The results of the tests led to these discoveries.