Generative Control of Time-Modulated Intelligent Reflecting Surfaces for Scalable Directional Modulation

Abstract

Time-modulated intelligent reflecting surfaces (TM-IRS) achieve directional modulation by switching their elements on and off at designed instants and durations, so that an undistorted signal reaches the intended user while appearing scrambled elsewhere. We propose a Flow Matching (FM)-based generative framework that designs these parameters for strong scrambling and reliable reception simultaneously. For orthogonal frequency-division multiplexing (OFDM) signals, we introduce a physics-informed reward that maximizes the time-modulation harmonic energy radiated in the null space of the legitimate-user response, subject to a minimum per-subcarrier signal-to-interference-plus-noise ratio (SINR), without requiring knowledge of adversary directions. Unlike prior generative flow network (GFlowNet) approaches, which assign discretized parameters sequentially, the FM model generates complete continuous configurations jointly, decoupling the generation depth from the number of elements and accelerating training. It also learns a distribution over configurations that can be resampled to vary the scrambling pattern. Numerical results show that, compared with a GFlowNet baseline, the proposed approach reduces the time to training loss near-saturation by $99.4\%$ on average for $3{\times}3$ to $12{\times}12$ surfaces, attains higher final rewards, scales to $80{\times}80$, and confines the low-symbol-error region more tightly around the legitimate user.