This work examines the use of two-way training in multiple-input multiple-output (MIMO) wireless systems to discriminate the channel estimation performances between a legitimate receiver (LR) and an unauthorized receiver (UR). This thesis extends upon the previously proposed discriminatory channel estimation (DCE) scheme that allows only the transmitter to send training signals. The goal of DCE is to minimize the channel estimation error at LR while requiring the channel estimation error at UR to remain beyond a certain level. If the training signal is sent only by the transmitter, the performance discrimination between LR and UR will be limited since the training signals help both receivers perform estimates of their downlink channels. In this work, we consider instead the two-way training methodology that allows both the transmitter and LR to send training signals. In this case, the training signal sent by LR helps the transmitter obtain knowledge of the transmitter-to-LR channel, but does not help UR estimate its downlink channel (i.e., the transmitter-to-UR channel). With transmitter knowledge of the estimated transmitter-to-LR channel, artificial noise (AN) can then be embedded in the null space of the transmitter-to-LR channel to disrupt UR's channel estimation without severely degrading the channel estimation at LR. Based on these ideas, two-way DCE training schemes are developed for both reciprocal and non-reciprocal channels. The optimal power allocation between training and AN signals is devised under both average and individual power constraints. Numerical results are provided to demonstrate the efficacy of the proposed two-way DCE training schemes.