Source-linked AI summary
Generative Control of Time-Modulated Intelligent Reflecting Surfaces for Scalable Directional Modulation
Kailong Wang, Sida Xiang, Athina Petropulu
TL;DR
Wireless directional modulation must preserve the intended user’s signal while scrambling reception elsewhere, but maximizing sum rate alone can suppress the harmonics needed for scrambling. The paper uses joint continuous Flow Matching control with a reward balancing null-space harmonic energy and a minimum per-subcarrier SINR. Compared with GFlowNet, it reaches training-loss near-saturation faster, produces stronger directional scrambling, and supports surfaces up to 80 × 80 elements.
Problem
Sum-rate maximization can drive continuous activation durations toward an always-on state that suppresses scrambling harmonics, while reliable reception must also be maintained.
Source: §2 · p. 2, §3.1 · p. 3Method
The paper jointly generates continuous TM-IRS timing configurations with Flow Matching and rewards null-space harmonic energy subject to a minimum per-subcarrier SINR gate.
Source: §1 · pp. 1–2, §3.1 · p. 3Results
Takeaways & Limitations
Resampling high-performing configurations can vary the spatial scrambling pattern, while joint updates decouple generation depth from the number of surface elements.
Source: §1 · pp. 1–2The hard SINR gate yields zero reward and no learning signal at 90 × 90 and 100 × 100, despite expected feasible configurations; addressing this is left for future work.
Source: §4 · p. 4
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
The introduction motivates directional modulation as a physical-layer security approach for scrambling wireless symbols outside authorized directions. It presents joint continuous Flow Matching control and a physics-informed reward as a scalable way to balance scrambling with legitimate-user reception.
1. INTRODUCTION
- Directional modulation preserves the desired signal constellation in authorized directions while intentionally scrambling symbols elsewhere.The paper situates it among physical-layer security methods that exploit channel and waveform properties.Source: §1 · p. 1
- TMA switching generates controllable harmonics that mix each OFDM subcarrier with symbols from all subcarriers, with mixing set by switch-on instants and durations.A rule-based design can eliminate intended-direction scrambling in the absence of noise.Source: §1 · p. 1
- Prior TM-IRS closed-form rules ignore noise and support only one communication-user direction, while GFlowNet approaches discretize and assign switching parameters sequentially.GFlowNet trajectory length grows with the number of control variables, and discretization limits the available timing values.Source: §1 · p. 1
- The FM controller jointly generates complete continuous switching configurations, avoiding sequential discretized decisions and decoupling generation depth from the number of IRS elements.Different Gaussian initializations can produce distinct high-performing configurations for varying the spatial scrambling pattern.Source: §1 · p. 1
- The physics-informed reward maximizes time-modulation harmonic energy in the legitimate-user response null space while enforcing a minimum per-subcarrier SINR, without requiring adversary directions.Source: §1 · p. 2
2. SYSTEM MODEL AND PROBLEM FORMULATION
The section formulates OFDM transmission through a time-modulated reflecting surface, where element timing shapes harmonic mixing and the received subcarrier symbols. It also identifies why maximizing sum rate alone can suppress the harmonics needed for scrambling.
2. SYSTEM MODEL AND PROBLEM FORMULATION
- The model uses an N-element ULA to transmit OFDM through an Mx × Mz TM-IRS, with K subcarriers and element-wise switch-on instants and activation durations.The IRS has M = MxMz reflecting elements; the incident OFDM waveform is defined using its subcarrier symbols and symbol duration.Source: §2 · p. 2
- Time-modulation harmonics mix subcarriers: harmonic l from subcarrier k contributes to received subcarrier i = k+l, where V0 scales the desired symbol and other coefficients produce intersubcarrier interference.Source: §2 · p. 2
- For the single-user setting, fixed IRS phase shifts cancel spatial phase at the legitimate user, leaving 2M timing variables to optimize.Source: §2 · p. 2
- When all activation durations approach one, the sum-rate reward approaches its upper bound while nonzero scrambling harmonics vanish, making the surface behave like an unswitched IRS.The discretized design in caps durations at (Q −1)/Q, leaving a gap of 1/Q from that endpoint.Source: §2 · p. 2
- The proposed reward addresses this degeneration by promoting nonzero-harmonic energy and assigning zero reward to configurations that fail the reception requirement.Source: §2 · p. 2
3. FLOW-MATCHING-BASED TM-IRS CONTROL
The FM controller transforms Gaussian samples into complete continuous TM-IRS timing configurations by jointly updating all timing parameters. Its physics-informed reward promotes harmonic energy in the legitimate-user null space while incorporating a minimum per-subcarrier SINR requirement.
3. FLOW-MATCHING-BASED TM-IRS CONTROL
- The controller starts from Gaussian noise, evolves states with a learned vector field, and maps the terminal state to a timing configuration.Sampling uses L Euler steps, with time progressing from initialization to completion.Source: §3 · pp. 2–3
- At each Euler step, FM jointly updates all 2M timing-state components, avoiding additional element-wise sequential decisions as the IRS grows.The terminal state is mapped through a sigmoid to a physical timing configuration in [0, 1)^2M.Source: §3 · p. 3
3.1. Physics-Informed Reward
- The sideband objective promotes nonzero-harmonic energy in the legitimate-user null space, where it does not contribute to that user’s received harmonic.The projected energy is formed using the effective steering vector for the transmitter–IRS–legitimate-user path.Source: §3.1 · p. 3
- Configurations violating the minimum SINR requirement receive zero reward, encouraging feasible designs even though individual generated samples may still violate it.Source: §3.1 · p. 3
- The FM policy generates groups of candidate configurations from Gaussian noise and updates using their rewards through Flow Policy Optimization.Each candidate is mapped to a physical timing configuration before its reward is evaluated.Source: §3.1 · p. 3
4. EXPERIMENTS AND RESULTS
Across the tested TM-IRS sizes, FM generally achieves higher final rewards than GFlowNet baselines and reaches near-saturation sooner. The proposed reward also produces stronger directional scrambling while preserving reception at the legitimate user, though the hard SINR gate limits feasibility at the largest tested sizes.
Experimental setup
- Evaluation uses an 8-antenna, 16-subcarrier OFDM system and measures SER with LMMSE detection over a one-degree angular grid.The reported setup uses QPSK, 0 dB SNR, and a 14 dB per-subcarrier SINR threshold.Source: §4 · p. 3
Directional scrambling
- On a 6 × 6 surface, the proposed reward yields high SER in unintended directions and low SER at the legitimate user, while GFlowNet has more low-SER regions elsewhere.With the sum-rate reward, 32 of 36 elements have switch-on durations above 0.99, suppressing nonzero harmonics and weakening scrambling.Source: §4 · p. 3, Figure 2 · p. 4
Final reward and network width
- From 3 × 3 onward, FM-Middle and FM-Large generally achieve higher rewards than GFN-Q4 and GFN-Q8, although FM-Middle leads FM-Large at 11 × 11 and 12 × 12.The comparison uses three independent seeds; at 2 × 2, FM models mostly failed to find feasible configurations.Source: §4 · p. 3, Figure 3 · p. 4
Training efficiency
- For 3 × 3 to 12 × 12 surfaces, near-saturation time falls by 99.4%, 98.8%, and 98.0% for FM-Small, FM-Middle, and FM-Large versus GFN-Q8, and by 99.4%, 98.8%, and 98.1% versus GFN-Q4.Near-saturation measures stabilization of each method’s training loss, not equality of final rewards.Source: §4 · pp. 3–4, Figure 3 · p. 4
SINR feasibility and scale
- From 4 × 4 to 80 × 80, both legitimate-user SINR pass rates are 100%; at 90 × 90 and 100 × 100, reward is zero and pass rates fall to 12.5% per subcarrier and 0% all-pass.The authors attribute the largest-size failure to the hard SINR gate: no feasible early samples leave the policy without a learning signal.Source: §4 · p. 4
5. CONCLUSION
The paper concludes that its Flow Matching framework jointly optimizes continuous TM-IRS timing parameters, reaches loss-based near-saturation faster than GFlowNet baselines, and supports arrays up to 80 × 80 elements.
5. CONCLUSION
- The framework reaches the loss-based near-saturation criterion faster than GFlowNet baselines and supports arrays up to 80 × 80 elements.The conclusion does not quantify the speedup.Source: §5 · p. 4
- Its target-null sideband-energy reward uses a minimum per-subcarrier SINR gate and assigns zero reward to configurations that violate the reception requirement.The desired signal phase is aligned through fixed IRS phase compensation.Source: §5 · p. 4
- Future work will extend the framework to multi-user scenarios by jointly optimizing element phase shifts and timing parameters, increasing the variables from 2M to 3M.Source: §5 · p. 4