GFlowNet-Based Antenna Selection for ISAC Systems Under the Presence of Eavesdroppers

This work addresses Physical Layer Security (PHY) in Integrated Sensing and Communication (ISAC) systems through antenna selection. We consider a MIMO transmitter communicating with a legitimate multi-antenna receiver while simultaneously sensing a target, with a multi-antenna eavesdrop-per intercepting the communication. Previous studies show that increasing the number of active antennas improves the ergodic secrecy rate up to a certain point, after which performance deteriorates. The instantaneous secrecy rate, for fixed channel conditions, depends not only on the number of active antennas but also on their positions. This work focuses on optimizing both the number and positions of active antennas to maximize the instantaneous secrecy rate, while ensuring that the Cramer Rao Bound (CRB) of the sensing functionality remains below a specified threshold. We propose an unsupervised approach based on Generative Flow Networks (GFlowNets), modeling antenna selection as a deterministic Markov Decision Process (MDP) where terminal states correspond to active antenna subsets of all sizes and configurations. GFlowNets parametrize the action-sampling distribution, ensuring the probability of reaching a final antenna subset configuration is proportional to its combination of the resulting secrecy rate and the CRB.