Security Risks in Vision-Based Beam Prediction: From Spatial Proxy Attacks to Feature Refinement

Avi Deb Raha, Kitae Kim, Mrityunjoy Gain, Apurba Adhikary, Zhu Han, Eui-Nam Huh, Choong Seon Hong

Researchers at Kyung Hee University and the University of Houston have identified a critical security gap in the next generation of wireless networks: the vulnerability of vision-assisted beamforming to adversarial attacks. Their work introduces a novel black-box attack and a lightweight defense mechanism that significantly improves both performance and robustness.

The Challenge: Securing 6G Beamforming

As wireless networks transition to extremely high frequencies (mmWave and Terahertz) for 6G, they require highly directional, "pencil-thin" beams to maintain signal integrity. In high-mobility environments like autonomous driving, base stations (BS) must constantly realign these beams with moving vehicles.

To reduce the massive overhead of traditional beam-sweeping, vision-based beam prediction has emerged as a solution. In this setup, cameras at the base station use deep learning to "see" a vehicle and predict the optimal beam index. However, deep learning models are notoriously fragile. While most adversarial research assumes a "white-box" scenario where the attacker knows model parameters, this paper addresses the more realistic black-box attack scenario where internal network data is inaccessible.

System model of vision-based beamforming under a Spatial Proxy Attack, where an adversary introduces perturbations to cause incorrect beam alignment.

Proposed Methodology

The researchers introduced two primary innovations to explore and mitigate these risks:

1. Spatial Proxy Attack (SPA)

The SPA is a clever black-box attack based on the realization that beam selection is essentially a "location-to-beam" mapping. An attacker does not need actual beam labels; instead, they segment the camera's field of view into horizontal "bins." By training a surrogate model to predict these bins, they can craft a perturbation that tricks the base station into perceiving the car in the wrong location, indirectly forcing a suboptimal beam choice.

2. Feature Refinement Module (FRM)

To counter such attacks, the authors designed the Feature Refinement Module (FRM). Unlike computationally expensive adversarial training, the FRM is a lightweight residual-based module. It learns to identify and subtract irrelevant features—such as background textures, weather noise, or adversarial perturbations—from the feature map before the final beam prediction is made.

Architecture of the proposed Feature Refinement Module (FRM) integrated with a ResNet backbone to filter out adversarial noise.

Experimental Setup and Datasets

The study utilized the DeepSense 6G dataset, a large-scale, real-world multi-modal dataset containing:

  • RGB Camera Images: Captured from a base station overlooking traffic.
  • mmWave Beam Power Measurements: Real-world signal data for 64 different beam directions.
  • Diverse Scenarios: Testing across five scenarios with varying traffic density, occlusion, and lighting.

Key Insights

  • Location as a Vulnerability: Spatial cues are the primary driver of vision-based beamforming. Manipulating the perceived location of a vehicle allows an attacker to hijack the process without knowing the underlying codebook.
  • Defense via Subtraction: Rather than adding complexity to make a model "stronger," the FRM assumes the initial feature map is "polluted" and refines it by removing noise.
  • The "Distillation" Trap: The researchers found that common techniques like Defensive Distillation perform poorly here. Because beam prediction relies on fine-grained spatial details, the "smoothing" effect of distillation blurs critical information, reducing robustness.

Results and Performance

The FRM-enhanced architecture was tested against standard backbones like ResNet-50 and MobileNetV2, yielding impressive results:

  • Clean Accuracy: A gain of up to +21.07% in Top-K accuracy under normal conditions.
  • Adversarial Robustness: A 41.31% increase in Top-1 robustness under attack compared to baseline models.
  • Noise Resilience: Up to a 47.66% improvement in performance when facing random Gaussian noise (simulating poor weather).
  • Efficiency: The module is lightweight and does not require explicit adversarial examples during training, making it ideal for real-time 6G applications.

Future Directions

While the FRM is highly effective, the authors noted that in cases where adversarial noise overlaps too closely with vehicle features, the module can accidentally suppress useful information. Future work will explore integrating uncertainty-aware or attention-based refinement to better distinguish between noise and task-relevant features, as well as testing against dynamic, adaptive beamforming codebooks.

Summary updated

BibTeX
@article{raha2026security,
  author = {Avi Deb Raha and Kitae Kim and Mrityunjoy Gain and Apurba Adhikary and Zhu Han and Eui-Nam Huh and Choong Seon Hong},
  title = {{Security Risks in Vision-Based Beam Prediction: From Spatial Proxy Attacks to Feature Refinement}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2026},
  doi = {10.1109/twc.2025.3634715},
}
Figure