Quantum Key Distribution (QKD) can be provably secure on paper, but in practice the security also depends on how the hardware is built and operated. A commercial QKD transmitter contains fast digital logic, analog driver electronics, and opto-electronic components that can radiate RF energy. If those emissions correlate with internal choices (such as state preparation settings) they may form a side channel that could be exploited at a distance (in a TEMPEST-like setting). Optical and protocol-level side channels in QKD have been studied extensively, but radiated RF leakage from commercial discrete-variable (DV) QKD equipment is less well understood. In this project we investigate whether weak, state-dependent RF signatures can be detected in a commercial DV-QKD system, and whether modern deep learning methods can reliably classify them under realistic measurement conditions. We will perform controlled near-field measurements of both magnetic (H-field) and electric (E-field) emissions using RF probes and a high-bandwidth oscilloscope. The transmitter can be set to repeatedly output fixed polarization states (e.g., H, V, and A), which provides a clean way to collect labeled training data where the intended quantum state is known. To avoid drawing conclusions from overly idealized data, we will record across multiple probe locations, orientations, and measurement sessions, capturing typical sources of variation such as placement tolerances, temperature drift, and ambient RF interference. Data will be split by session (rather than by short time segments) to ensure a realistic separation between training, validation, and test sets. From the recorded waveforms we will compute time–frequency features such as short-time Fourier transform spectrograms, and we will also examine components tied to synchronous device activity (for example clock- and symbol-rate harmonics). We will train deep learning models (primarily convolutional networks and sequence models) to classify emissions collected in fixed-state mode, and we will use model interpretation and ablation studies to identify which frequency bands and time windows carry the most discriminative information. We will then apply the trained models to measurements taken during normal QKD operation with randomized basis and bit selection, to test whether any learned leakage persists when the system runs as intended and additional internal activity is present. Where synchronized metadata or internal logs are available we will quantify classification performance and estimate information leakage; otherwise we will test for statistically significant deviations from chance and evaluate stability across sessions and probe placements. The project is done by LiU (ISY, Division of Information Coding), together with an industry partner.