Robust Adversarial Attacks Against DNN-Based Wireless Communication Systems

Alireza Bahramali, Milad Nasr, Amir Houmansadr, Dennis Goeckel, Don Towsley

Introduction

Deep Neural Networks (DNNs) have become integral to modern wireless communication systems, facilitating tasks such as signal detection, channel estimation, and modulation recognition. However, similar to their vulnerabilities in computer vision, wireless DNNs possess "blind spots" susceptible to adversarial examples. In the paper "Robust Adversarial Attacks Against DNN-Based Wireless Communication Systems," researchers from UMass Amherst demonstrate that these systems are remarkably fragile. They introduce a novel framework capable of generating adversarial attacks that are both effective and resilient against detection and filtering.

The Challenge of Wireless Adversarial Attacks

Attacking a wireless system presents unique difficulties compared to digital image domains:

  1. The "A-Priori" Problem: An attacker typically lacks knowledge of the exact signal being transmitted at any given moment. This necessitates an input-agnostic attack—a universal perturbation that degrades performance regardless of the underlying data.
  2. The Detection Problem: Wireless channels are inherently noisy due to Additive White Gaussian Noise (AWGN). If adversarial noise is static or statistically distinct from natural noise, a sophisticated receiver can estimate and subtract it, or flag the transmission as under attack.

Existing attacks often rely on a single, static noise vector, making them easy targets for modern defensive filters.

Methods: The Perturbation Generator Model (PGM)

To overcome these limitations, the authors developed the Perturbation Generator Model (PGM). Instead of optimizing for a single noise vector, the PGM is a neural network trained to generate a diverse library of Universal Adversarial Perturbations (UAPs).

The PGM architecture generates diverse, undetectable perturbations by mimicking natural channel noise and employing distance regularization.

The framework incorporates several key innovations:

  • Stealth via GANs: To ensure undetectability, the PGM utilizes a Generative Adversarial Network (GAN) structure. A discriminator attempts to distinguish between the generated adversarial noise and standard AWGN. The PGM learns to produce noise that is statistically indistinguishable from natural interference.
  • Diversity for Robustness: A distance regularizer is employed to force the PGM to generate UAPs that are mathematically distinct from one another. This prevents a defender from "averaging out" the attack; even if one perturbation is identified, the next will be fundamentally different.
  • Black-Box Transferability: The attacks remain potent even when the attacker does not know the target's exact DNN architecture. By training against a substitute model, the generated perturbations successfully transfer to the actual target system.

Experimental Evaluation

The PGM attack was evaluated across three critical wireless applications: end-to-end autoencoders, modulation recognition (using the RML2016.10a dataset), and OFDM signal detection.

Key Results

The quantitative results demonstrate the devastating impact of the PGM attack across different systems compared to baseline "single UAP" attacks.

Performance degradation across (a) Autoencoder communication systems, (b) Modulation recognition systems, and (c) OFDM systems under PGM and single UAP attacks.
  • Autoencoders: At a perturbation-to-signal ratio (PSR) of -10dB, the PGM attack increases the Block-Error Rate (BLER) significantly more than traditional single-vector UAP attacks.
  • Modulation Recognition: Under active defense scenarios, the PGM attack reduced classification accuracy from 0.69 to approximately 0.23, whereas older attacks were largely mitigated.
  • OFDM Systems: The attack resulted in a 9x increase in Bit Error Rate (BER), effectively disrupting the communication link.

Conclusion

The research highlights a critical structural vulnerability in DNN-based wireless systems. By treating adversarial perturbations as a generative task rather than a static optimization problem, the PGM creates a "moving target" that bypasses conventional defenses. Future work will focus on the challenges of phase synchronization in high-speed mobile environments and the deployment of these attacks using Software Defined Radios (SDRs) to test resilience against real-world hardware impairments.

Summary updated

BibTeX
@inproceedings{bahramali2021robust,
  author = {Alireza Bahramali and Milad Nasr and Amir Houmansadr and Dennis Goeckel and Don Towsley},
  title = {{Robust Adversarial Attacks Against DNN-Based Wireless Communication Systems}},
  booktitle = {CCS},
  year = {2021},
  doi = {10.1145/3460120.3484777},
}

Loading related content...

Figure