Defending Adversarial Attacks on Deep Learning-Based Power Allocation in Massive MIMO Using Denoising Autoencoders

Rajeev Sahay, Minjun Zhang, David J. Love, Christopher G. BrintonIEEE Trans. Cogn. Commun. Netw 2023 · OpenReview

TL;DR

This paper introduces a novel defense mechanism using Denoising Autoencoders (DAEs) to protect deep learning models for power allocation in massive MIMO (maMIMO) networks from adversarial attacks. The DAE effectively filters out malicious perturbations from input data, ensuring reliable power allocation even when under attack, outperforming existing defense strategies.

What Problem Does This Paper Address?

Massive MIMO (maMIMO) systems, with their vast number of antennas, offer significant improvements in wireless communication efficiency. A critical task in these systems is power allocation, determining how much power to distribute to each user (UE) to maximize performance while staying within the base station's (BS) power budget. Deep learning models, particularly Deep Neural Networks (DNNs), have emerged as a powerful tool for achieving near-optimal power allocation with reduced computational complexity compared to traditional optimization methods.

However, these DNNs are vulnerable to "adversarial attacks." In this context, an adversarial attack involves subtly altering the input data (in this case, UE positions) in a way that's imperceptible to humans but causes the DNN to produce an incorrect, often infeasible, power allocation. This can lead to unfair resource distribution and degrade the overall network performance. Existing defenses often require retraining the DNN or suffer from performance degradation on clean data, creating a need for a robust and efficient defense.

Methods

The core of this paper's contribution is the Denoising Autoencoder (DAE) defense. The DAE is a neural network trained to reconstruct clean data from corrupted inputs. Here's how it works:

  1. Training the DAE: The DAE is trained to map potentially perturbed input data (UE positions) back to their original, unperturbed state. Crucially, it's also trained to map clean data to itself, ensuring that normal operation isn't negatively impacted. The training process specifically uses adversarial examples generated by the Projected Gradient Descent (PGD) attack, a potent method for crafting adversarial perturbations.
  2. Inference with the DAE: During real-time operation, when a BS receives UE position data, it's first passed through the trained DAE. The DAE "denoises" the input, removing any adversarial perturbations.
  3. Power Allocation with DNN: The denoised input is then fed into the original DNN responsible for power allocation. Because the input is now clean, the DNN can produce a reliable and feasible power allocation.

This approach is advantageous because it acts as a pre-processing step, meaning the original power allocation DNN doesn't need to be retrained.

Datasets

The authors utilized a publicly available dataset: Power Allocation in Multi-Cell Massive MIMO dataset. This dataset was used for both training the DAE and evaluating the performance of the proposed defense. The evaluation involved simulating a system with L=4 cells in a 2x2 grid layout.

Key Insights and Ideas

  • DAEs as a General Defense for Regression DNNs: While DAEs have shown success in computer vision for image classification, this work demonstrates their effectiveness in a regression task (power allocation) within wireless communications.
  • Robustness to Different Attack Types: The DAE, trained on PGD-generated attacks, proved effective against other first-order l∞-bounded adversarial attacks like FGSM, highlighting its generalization capabilities.
  • Pre-processing is Viable in Wireless: Unlike in computer vision where pre-processing defenses are sometimes dismissed due to full adversary knowledge, in wireless, adversaries often have limited information, making pre-processing a practical and effective defense.
  • Balancing Performance and Robustness: The DAE defense significantly mitigates adversarial attacks without substantially degrading performance on clean, unperturbed data, a common issue with other defense methods like adversarial training.

Key Results

  • Significant Attack Mitigation: The DAE defense dramatically reduced the attack success rate (percentage of infeasible power allocations) across various attack types (FGSM, PGD), precoding schemes (MMMSE, MR), and threat models (semi-white box, black box). For instance, in a semi-white box PGD attack scenario, the DAE reduced infeasible solutions from nearly 100% to approximately 10%.
  • Outperforms Baselines: The DAE consistently outperformed existing defense methods like adversarial training and defensive distillation, especially against higher-bounded perturbations.
  • Minimal Performance Degradation on Clean Data: In the absence of attacks, the DAE defense maintained performance comparable to the original DNN, with slight improvements observed in some cases.
  • Low Computational Overhead: The DAE adds a minimal computational cost during online deployment, with forward propagation taking only an additional 0.87 ms, allowing for rapid power allocation updates even with UEs in motion. Offline training of the DAE was also found to be computationally efficient.

Limitations and Future Work

  • Threat Model Assumption: The defense is primarily evaluated against semi-white box and black box threat models, which are considered more realistic for wireless communication scenarios where adversaries have limited knowledge. The effectiveness against a full white-box attack (where the adversary has complete knowledge of both the DNN and the DAE) was not explicitly tested.
  • Scalability to Larger Networks: While the paper demonstrates effectiveness in a 4-cell system, future work could explore the scalability of this approach to much larger, federated machine learning networks.
  • Exploration of Other Attack Types: While robust against l∞-bounded attacks, further investigation into defenses against other types of adversarial perturbations could be beneficial.

Summary updated

Abstract

Recent work has advocated for the use of deep learning to perform power allocation in the downlink of massive MIMO (maMIMO) networks. Yet, such deep learning models are vulnerable to adversarial attacks. In the context of maMIMO power allocation, adversarial attacks refer to the injection of subtle perturbations into the deep learning model’s input, during inference (i.e., the adversarial perturbation is injected into inputs during deployment after the model has been trained) that are specifically crafted to force the trained regression model to output an infeasible power allocation solution. In this work, we develop an autoencoder-based mitigation technique, which allows deep learning-based power allocation models to operate in the presence of adversaries without requiring retraining. Specifically, we develop a denoising autoencoder (DAE), which learns a mapping between potentially perturbed data and its corresponding unperturbed input. We test our defense across multiple attacks and in multiple threat models and demonstrate its ability to (i) mitigate the effects of adversarial attacks on power allocation networks using two common precoding schemes, (ii) outperform previously proposed benchmarks for mitigating regression-based adversarial attacks on maMIMO networks, (iii) retain accurate performance in the absence of an attack, and (iv) operate with low computational overhead. Code is publicly available at https://github.com/Jess-jpg-txt/DAE_for_adv_attacks_in_MIMO.
BibTeX
@article{sahay2023defending,
  author = {Rajeev Sahay and Minjun Zhang and David J. Love and Christopher G. Brinton},
  title = {{Defending Adversarial Attacks on Deep Learning-Based Power Allocation in Massive MIMO Using Denoising Autoencoders}},
  journal = {IEEE Trans. Cogn. Commun. Netw.},
  year = {2023},
  url = {https://openreview.net/forum?id=qYNB1ZtpFQ},
}
Figure