Deep Learning-Driven Friendly Jamming for Secure ISAC Under Channel Uncertainty

Integrated sensing and communication enables efficient spectrum sharing by jointly performing radar sensing and wireless transmission. This paper proposes a deep learning-driven framework for secure ISAC under imperfect channel state information (CSI) and unknown eavesdropper (Eve) locations. Unlike conventional friendly jamming (FJ) methods requiring Eve's CSI or precise angle-of-arrival (AoA), the proposed approach leverages radar echo feedback to guide directional jamming without explicit Eve information. To ensure robustness against sensing uncertainty, we introduce a radar-aware neural network that jointly optimises beamforming and jamming using a nonparametric Fisher Information Matrix (FIM) estimator based on f-divergence. The design satisfies Cramér-Rao lower bound constraints under noisy AoA. A quantised tensor-train encoder reduces model size substantially while maintaining performance. A non-overlapping secure subband scheme further isolates communication channels. Simulations demonstrate higher secrecy rates, reliable BLER, and robustness to CSI and angular errors, confirming the practicality of learning-based secure ISAC.