This dataset is generated with QuaDRiGa under the 3GPP_38.901_Indoor channel model, considering both LOS and NLOS propagation conditions. The scenario places a fixed base station (Alice) and a mobile user (Bob) on a random indoor linear trajectory, with a passive eavesdropper (Eve) positioned at a fixed separation from Bob. The system operates at 95 GHz with a TDD configuration, where the uplink and downlink channel observations are separated by a fixed delay of 0.1 ms to emulate the reciprocity impairment of a half-duplex TDD system. Channel snapshots along the trajectory are mapped to the frequency domain using an OFDM grid of 128 subcarriers with 60 kHz subcarrier spacing (7.68 MHz total bandwidth).
Each .mat file contains h_Alice, h_Bob and h_Eve, complex frequency-domain channel coefficients of shape [subcarriers × snapshots], representing the estimated CFR at each node after pilot-based OFDM channel estimation. The reciprocity mismatch between h_Alice and h_Bob is the target that the AI model is trained to compensate, while h_Eve is retained for security evaluation of the resulting PKG scheme.