Adversarial Attacks on Graph Neural Networks Based Spatial Resource Management in P2P Wireless Communications

Ahmad Ghasemi, Ehsan Zeraatkar, Majid Moradikia, Seyed A. ZekavatIEEE Trans. Vehicular Technology 2024 · OpenReview

TL;DR

This paper introduces novel adversarial attacks against Graph Neural Networks (GNNs) used for resource management in P2P wireless networks. The research demonstrates that these attacks can significantly degrade communication quality and alter the statistical properties of the network's channel eigenvalues, highlighting a critical vulnerability in these systems.

What Problem Does This Paper Address?

Machine learning, particularly Graph Neural Networks (GNNs), has shown great promise in optimizing complex communication systems like radio resource management. However, like many AI models, GNNs are susceptible to adversarial attacks. These attacks aim to manipulate the GNN's behavior by subtly altering its inputs, leading to incorrect decisions and degraded performance. While adversarial attacks have been studied in other domains like computer vision, their impact on GNNs specifically within the context of P2P wireless communications, where resource allocation is crucial for efficiency and reliability, remains less explored. This paper addresses this gap by developing and analyzing adversarial attacks targeting GNN-based spatial resource management in P2P wireless networks, and also proposes a method for detecting these attacks.

Methods

The core of this work involves developing four new adversarial attacks designed to degrade the performance of a GNN-based radio resource management system in P2P wireless communications. These attacks operate during the testing phase, meaning the GNN model itself is already trained. The attacks are categorized based on whether they target the "vertices" (representing transceiver pairs) or "edges" (representing interference channels) of the graph model.

Key methodological aspects include:

  • Attack Formulation: The attacks are framed as optimization problems where the adversary aims to minimize the system's Quality of Communication (QoC), often measured as a weighted sum rate.
  • Constraints: The attacks consider realistic constraints on the adversary, such as:
  • Channel-bounded (B): Limiting the number of channels that can be perturbed.
  • Power-bounded (Bp): Limiting the adversary's transmission power.
  • Min-detectable: Crafting attacks to minimize the probability of being detected by the system.
  • Heuristic Algorithms: Due to the non-convex nature of the optimization problems, heuristic algorithms are proposed to find near-optimal solutions for the attack strategies.
  • Attack Detection: A novel Min-Max optimization problem based on the One-Sample Kolmogorov-Smirnov (KS) Test is proposed for detecting these attacks. This method analyzes the distribution of channel tensor eigenvalues before and after an attack.

Datasets

The paper does not explicitly mention pre-existing datasets for training or evaluation. Instead, the authors generated their own simulation data. The simulation setup involved:

  • Network Generation: Creating wireless networks with a varying number of transceiver pairs (N) distributed within a defined rectangular area.
  • Channel Model: Utilizing a standard channel model that accounts for path loss, antenna gain, shadowing, and small-scale fading.
  • Data Split: A split of 2000 samples for training and 500 samples for testing was used.
  • Scenarios: Various scenarios were simulated to evaluate the attacks, including different numbers of users (N), signal-to-noise ratios (SNR), and TX-RX distances.

Key Insights and Ideas

  • GNNs are Vulnerable in Wireless: The research clearly demonstrates that GNNs used for critical functions like resource management in wireless networks are not immune to adversarial manipulation.
  • Edge Attacks are More Potent: Perturbing interference channels (edges) generally leads to a more severe degradation of communication quality compared to perturbing transceiver pairs (vertices). This is because interference channels impact multiple users simultaneously.
  • Detection via Eigenvalue Distribution: Analyzing the statistical distribution of channel tensor eigenvalues is a viable method for detecting adversarial attacks. Attacks significantly alter this distribution, sometimes even changing its fundamental type.
  • Trade-off Between Effectiveness and Detectability: Adversaries face a trade-off: more destructive attacks are often easier to detect. The proposed attacks attempt to balance these factors using constraints like power limitations and minimizing detectability.
  • Johnson's SU Distribution is Key: In the absence of attacks, the channel eigenvalues consistently follow the Johnson's SU distribution. This provides a baseline for detecting deviations caused by adversarial actions.

Key Results

  • Significant QoC Degradation: The proposed adversarial attacks can drastically reduce the total Quality of Communication (QoC). The most effective attack, the "Bc edge perturbation," achieved an approximate 95% reduction in total QoC.
  • Impact of Network Size and Distance: The effectiveness of attacks generally increases with the number of users (N). Edge perturbations are particularly effective in homogeneous networks (consistent TX-RX distances), while vertex perturbations are more damaging in heterogeneous networks.
  • SNR Independence: The performance of the adversarial attacks was found to be largely independent of the Signal-to-Noise Ratio (SNR).
  • Eigenvalue Distribution Alteration: Adversarial attacks, particularly the "Bc edge perturbation," were shown to alter the type of the eigenvalue distribution, moving away from the baseline Johnson's SU distribution. Other attacks primarily modified the parameters of this distribution.
  • Detection Efficacy: The proposed KS-test based detection method successfully identified deviations in eigenvalue distributions caused by attacks.

Limitations and Future Work

The authors acknowledge several limitations that pave the way for future research:

  • Adversary-to-CPU Channel: The current model assumes a perfect channel between the adversary and the Central Processing Unit (CPU). In reality, this channel might be noisy or unreliable, potentially limiting attack effectiveness.
  • Perfect CSI Knowledge: The adversary is assumed to have perfect Channel State Information (CSI). In practice, CSI can be inaccurate, which might reduce the success rate of attacks.
  • Single Adversary Assumption: The study focuses on a single adversary. Future work could explore scenarios with multiple adversaries, which could lead to more complex and coordinated attacks.
  • Controlled Environments: The assumptions made are more applicable to static and controlled wireless environments. Dynamic environments with changing conditions and multiple adversaries present more challenging scenarios for both attacks and defenses.
  • Distributed Deployment: The current work focuses on centralized GNNs. Extending the analysis to distributed GNN architectures in wireless systems is a promising future direction.

Summary updated

Abstract

This article introduces adversarial attacks targeting a Graph Neural Network (GNN)-based radio resource management system in point-to-point (P2P) communications. Our focus lies on perturbing the trained GNN model during the test phase, specifically targeting its vertices and edges. To achieve this, four distinct adversarial attacks are proposed, each accounting for different constraints, aiming to manipulate the behavior of the system. The proposed adversarial attacks are formulated as optimization problems, aiming to minimize the system's communication quality. The efficacy of these attacks is investigated against the number of users, signal-to-noise ratio (SNR), and adversary power budget. Furthermore, we address the detection of such attacks from the perspective of the Central Processing Unit (CPU) of the system. To this end, we formulate an optimization problem that involves analyzing the distribution of channel eigenvalues before and after the attacks are applied. This formulation results in a Min-Max optimization problem, allowing us to detect the presence of attacks. Through extensive simulations, we observe that in the absence of adversarial attacks, the eigenvalues conform to Johnson's SU distribution. However, the attacks significantly alter the characteristics of the eigenvalue distribution, and in the most effective attack, they even change the type of the eigenvalue distribution.
BibTeX
@article{ghasemi2024adversarial,
  author = {Ahmad Ghasemi and Ehsan Zeraatkar and Majid Moradikia and Seyed A. Zekavat},
  title = {{Adversarial Attacks on Graph Neural Networks Based Spatial Resource Management in P2P Wireless Communications}},
  journal = {IEEE Trans. Vehicular Technology},
  year = {2024},
  url = {https://openreview.net/forum?id=P4WS3Wemvy},
}
Figure