Causal AI-Driven Secure mmWave Beam Alignment with Defensive Variational Autoencoders

Nasir Khan, Asmaa Abdallah, Abdulkadir Celik, Ahmed M. Eltawil, Sinem Coleri, Cicek CavdararXiv 2026 · arXiv

In the race toward 6G, millimeter-wave (mmWave) frequencies are the prize, offering the massive bandwidth needed for ultra-high data rates. But there’s a catch: these signals are fragile. To maintain a link, base stations must use highly directional "narrow beams," which require precise alignment. Traditional methods for finding the right beam are either painfully slow or vulnerable to "adversarial attacks"—malicious noise designed to trick AI-based controllers.

A new paper, "Causal AI-Driven Secure mmWave Beam Alignment with Defensive Variational Autoencoders," proposes a clever two-pronged solution. By combining causal discovery with generative AI, the authors have created a framework that is not only significantly faster than current deep learning (DL) approaches but also remarkably resilient to security threats.

TL;DR

The authors propose a framework that uses Causal AI to identify the minimal set of signal measurements needed for mmWave beam alignment, reducing overhead by up to 89.8%. They further introduce a Defensive Variational Autoencoder (DVAE) that "denoises" adversarial attacks, maintaining high accuracy even when the system is under active interference.

What Problem Does This Paper Address?

In mmWave systems, the base station (BS) and user equipment (UE) must find the best pair of beams to communicate.

  1. The Overhead Problem: Standard "exhaustive search" sweeps through every possible beam, which is slow and energy-intensive. DL-based methods speed this up by predicting the best beam from a few measurements, but they often rely on "spurious correlations"—patterns in the data that aren't actually related to the physics of the channel. This leads to poor generalization in new environments.
  2. The Security Problem: DL models are notoriously "brittle." An attacker can inject tiny, calculated perturbations into the received signal strength indicators (RSSI) to mislead the beam classifier, causing the link to drop.
Figure 1: System model illustration for the beam alignment in mmWave communication system.

Methods

The proposed framework operates in two main stages:

1. Causal Feature Selection

Instead of feeding all possible signal measurements into a neural network, the authors use DirectLiNGAM (a causal discovery algorithm) to build a Directed Acyclic Graph (DAG). This graph identifies which specific measurements actually cause the optimal beam selection. By selecting only these "causal parents," the system ignores redundant or misleading data.

2. Defensive Variational Autoencoder (DVAE)

To handle security, the authors place a DVAE in front of the beam classifier. The DVAE is trained to take potentially "attacked" (noisy) inputs and map them back to their "clean" versions.

  • Encoder: Compresses the input into a lower-dimensional latent space.
  • Decoder: Reconstructs the clean signal from that latent representation. The entire pipeline is then fine-tuned end-to-end so the classifier learns to work specifically with these denoised representations.
Figure 2: Illustration of the proposed causal feature selection and DVAE-assisted robust beam alignment framework, where the BS hosts both the DNN beam classifier and the pre-processing DVAE to mitigate potential adversarial attacks.

Datasets

The study utilized the DeepMIMO dataset, specifically the Boston5G scenario. This is a high-fidelity ray-tracing simulation of downtown Boston, providing realistic channel vectors for a 32-antenna base station and hundreds of single-antenna users. The dataset includes over 57,000 samples, split into training, validation, and testing sets.

Key Insights and Ideas

  • Causality vs. Correlation: Most AI in wireless just looks for patterns. This paper argues that by finding causal links, we can achieve better performance with much less data. The resulting feature set closely matches those found by complex "Explainable AI" (XAI) methods like SHAP but at a fraction of the cost.
  • Efficiency of Causal Discovery: The DirectLiNGAM approach was found to be 829 times faster per sample than SHAP for feature selection. This makes it feasible for the rapidly changing environments of 6G.
  • Generative Defense: Unlike "defensive distillation" (a common technique that smooths out model decision boundaries), the DVAE actually fixes the corrupted data. This makes it much harder for an attacker to find a "blind spot" in the model.
Figure 3: DAG obtained via DirectLiNGAM based causal beam selection approach.

Key Results

  • Overhead Reduction: The causal approach reduced beam-sweeping overhead by 59.4% compared to standard DL methods and 89.8% compared to exhaustive search.
  • Robustness: Under adversarial attacks (both "semi-white box" and "black box"), the DVAE-based defense sustained 85% accuracy.
  • Performance Gain: The DVAE outperformed existing defensive distillation schemes by 15–20% on average across various attack scenarios.
  • Inference Speed: The total inference time (DVAE + Classifier) is roughly 3.48 milliseconds, well within the requirements for real-time 5G/6G operation.

Limitations and Future Work

The authors acknowledge that their current model was tested in a relatively static urban simulation. Real-world 6G environments will involve high mobility (e.g., vehicles and drones). Future work will focus on:

  • Adapting the causal AI framework to time-varying channels.
  • Testing the system's performance in highly mobile scenarios where beam tracking becomes as important as initial alignment.
  • Exploring how these causal models can generalize across entirely different cities or environments without full retraining.

Summary updated

Abstract

Efficient and secure beam alignment is essential for mmWave communication systems. However, existing deep learning (DL)-based beam alignment methods often neglect causal relationships between inputs and outputs, resulting in limited interpretability, poor generalization, and excessive beam sweeping. Moreover, these DL solutions are vulnerable to adversarial attacks, where small perturbations in received power measurements can mislead beam classifiers. To overcome these issues, we propose a causally-aware DL framework integrating causal discovery into beam alignment. A two-stage causal beam selection algorithm learns a Bayesian graph capturing dependencies between received signal strength indicators (RSSI) as inputs and the optimal beam index as the target variable, then uses it to select a minimal set of relevant input features for the DL classifier. To mitigate adversarial attacks, we introduce a defensive variational autoencoder (DVAE) that maps potentially perturbed inputs back to their clean versions. End-to-end fine-tuning of the downstream beam classifier on the DVAE-reconstructed RSSI representations further boosts prediction accuracy. Simulation results show that causal beam selection achieves accuracy comparable to conventional methods while providing an \(\mathbf{829\times}\) lower per-sample computational cost than SHAP (Shapley Additive Explanations), a widely used explainable AI (XAI)-based feature selection method. It further reduces beam-sweeping overhead by \(\mathbf{59.4\%}\) and \(\mathbf{89.8\%}\), compared to conventional DL-based beam selection and exhaustive search, respectively. The DVAE-based defense effectively denoises adversarial inputs, sustaining 85\% accuracy across different perturbation bounds, retains baseline performance on unperturbed RSSI inputs and surpasses defensive distillation defense scheme by 15--20\% on average under different attack scenarios.
BibTeX
@misc{khan2026causal,
  author = {Nasir Khan and Asmaa Abdallah and Abdulkadir Celik and Ahmed M. Eltawil and Sinem Coleri and Cicek Cavdar},
  title = {{Causal AI-Driven Secure mmWave Beam Alignment with Defensive Variational Autoencoders}},
  year = {2026},
  eprint = {2611.18321},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18321},
}
Figure