TL;DR
Vision-based beam prediction for 6G mmWave communications is highly efficient but dangerously vulnerable to visual manipulation. This paper introduces the first black-box attack (Spatial Proxy Attack) targeting these systems by exploiting spatial cues, and counters it with a lightweight Feature Refinement Module (FRM) that strips away adversarial noise to boost both clean accuracy and network robustness.
What Problem Does This Paper Address?
Next-generation 6G networks rely heavily on millimeter-wave (mmWave) and terahertz (THz) frequencies. While these bands offer massive data rates, they require highly directional, narrow beams. In fast-moving vehicular environments, constantly adjusting these beams creates massive computational overhead. To solve this, engineers use RGB cameras mounted on base stations to visually track vehicles and predict the optimal beam direction.
However, deep learning models are notoriously fragile. A malicious actor could subtly perturb the camera feed to trick the system into choosing a suboptimal beam, degrading the Quality of Service (QoS). Traditional "white-box" attacks assume the attacker knows the internal model parameters or exact beam labels, which is unrealistic in real-world telecom deployments. This paper addresses two major gaps:
- How can an adversary effectively launch a realistic, real-time black-box attack without access to proprietary network labels or model architectures?
- How can we protect these vision-based beamforming systems from both malicious attacks and random environmental noise without sacrificing performance on clean data?
Methods
The Attack: Spatial Proxy Attack (SPA)
Because a base station's codebook maps specific beams to specific physical sectors, vision-based beamforming models inherently learn a "location-to-beam" mapping. The authors exploit this via the Spatial Proxy Attack (SPA). Instead of guessing exact beam indices, the attacker segments the camera's image plane into horizontal bins (spatial proxies). Using an off-the-shelf object detector (YOLOv8), the attacker labels which bin a vehicle is in and trains a surrogate model on this proxy task.
To make the attack practical for real-time streaming, they generate a Universal Adversarial Perturbation (UAP) offline using a Fast Gradient Sign Method (FGSM) variant. This single, static noise pattern can be stamped onto every incoming camera frame, consistently tricking the base station into pointing its beam away from the target vehicle.
The Defense: Feature Refinement Module (FRM)
Instead of relying on computationally expensive adversarial training or defensive distillation (which often degrades clean accuracy), the authors introduce a hybrid deep learning architecture featuring a lightweight Feature Refinement Module (FRM).
Placed right after a standard convolutional backbone (like ResNet-50 or MobileNetV2), the FRM uses a residual-based subtraction strategy. It explicitly learns to identify background clutter, random noise, and adversarial patterns within the high-dimensional feature map and subtracts them:
$$F_{\text{refined}} = F_{\text{backbone}} - \text{FRM}(F_{\text{backbone}}; \psi)$$
This leaves a clean, highly focused representation containing only the essential spatial cues needed for accurate beam prediction.
Datasets
The framework was evaluated using the DeepSense 6G dataset (specifically Scenarios 1, 2, 3, 4, and 7). This dataset contains real-world multi-modal measurements collected from a practical mmWave communication testbed operating at the 60 GHz frequency band. The visual data consists of $960 \times 540$ RGB camera images paired with ground-truth received power values for 64 predefined transmit beams, framing the task as a 64-class classification problem. The scenarios cover diverse environments like urban intersections, highways, and parking lots with varying levels of mobility, multi-vehicle interference, and lighting conditions.
Key Insights and Ideas
- Location as a Vulnerability Proxy: The core breakthrough of SPA is realizing that you don't need to know a network's proprietary beamforming codebook to disrupt it. Tricking the model about where the car is located automatically forces it to select the wrong beam.
- The Failure of Defensive Distillation: Traditional deep learning defenses like defensive distillation actually perform poorly in beam prediction. Distillation smooths out decision boundaries to help generalization, but beam selection relies on highly localized spatial features (the exact position of a car). Smoothing these boundaries dilutes the fine-grained spatial precision required for accurate targeting.
- Robustness via Subtraction: Rather than forcing a network to tolerate noise, the FRM succeeds by actively stripping it away. This joint optimization ensures that the model performs better not just under attack, but also under normal, clean operating conditions.
Key Results
- Clean Accuracy Gains: Integrating the FRM consistently outperformed standard baselines, achieving up to a +21.07% gain in Top-K accuracy under clean conditions.
- Adversarial Robustness: Under SPA attacks, the FRM-enhanced architecture delivered up to a 41.31% increase in Top-1 adversarial robustness and up to a +61.04% robustness enhancement overall compared to baseline models.
- Noise Resilience: When subjected to random Gaussian environmental noise, the proposed module achieved up to a 47.66% improvement in accuracy, proving its real-world utility against non-malicious camera degradation.
Limitations and Future Work
The authors note that the FRM is not entirely infallible. In specific edge cases where adversarial noise heavily overlaps with salient vehicle features (such as in highly cluttered multi-vehicle scenarios), the module's refinement strategy can mistakenly suppress useful spatial information, leading to slight drops in performance compared to unrefined baselines.
To address this, the authors suggest that future work should explore attention-based or uncertainty-aware refinement mechanisms to dynamically separate malicious noise from task-relevant features under extreme environmental conditions.
Summary updated