AI-Native Task-Oriented Semantic Communications over a Relay: Benefits and Vulnerabilities

Yalin E. Sagduyu, Tugba Erpek, Aylin YenerarXiv 2026 · arXiv

TL;DR

This paper introduces a joint end-to-end framework for relay-assisted semantic communication (SemCom) that significantly boosts task accuracy and image reconstruction in noisy wireless environments. However, the authors also reveal a critical vulnerability: a compromised "Byzantine" relay can use a lightweight neural network to inject structured perturbations that effectively "blind" the destination's AI models, reducing accuracy by nearly 90%.

What Problem Does This Paper Address?

Traditional wireless communication focuses on bit-level fidelity—ensuring the 0s and 1s sent are the same as the ones received. Semantic Communication (SemCom) shifts this paradigm by focusing on the "meaning" or the specific task at hand (e.g., classifying an image) rather than perfect data reconstruction. This is much more efficient in bandwidth-constrained environments.

While most SemCom research focuses on direct source-to-destination links, real-world wireless systems often rely on relays to bridge long distances or overcome poor channel conditions. This paper tackles two main questions:

  1. How do we jointly design the AI models at the source, relay, and destination to maximize performance?
  2. Does introducing an AI-powered relay create new security risks? Specifically, what happens if the relay itself is an "insider threat" (a Byzantine adversary)?

Methods

The authors propose a system where a source node encodes an image into a compact latent representation. This signal is broadcast to both a relay and the destination. The relay then processes the signal and forwards it to the destination in a second phase.

Fig. 1: System model of semantic relay communications.

The paper compares three relaying strategies:

  • Amplify-and-Forward (AF): A simple baseline that just normalizes and retransmits the received signal.
  • Decode-and-Forward (DF): The relay attempts to reconstruct the full image and then re-encodes it for transmission.
  • Learn-and-Forward (LF): The authors' novel approach, which uses a lightweight 1D-Convolutional Neural Network (CNN) to perform a direct latent-to-latent transformation without ever reconstructing the original image.

To test vulnerabilities, they formulate a Byzantine attack. The adversarial relay uses a trained "Perturbation Generator" (another lightweight DNN) to inject structured noise into the forwarded signal. This attack is "untargeted," meaning it aims to maximize the destination's uncertainty (entropy) and drive its predictions away from the correct ones using a Kullback–Leibler (KL) divergence loss.

Fig. 3: Attack on semantic relay communications.

Datasets

The researchers evaluated their models using the CIFAR-10 dataset, a standard benchmark consisting of 60,000 32x32 color images across 10 classes (e.g., airplanes, dogs, trucks).

Key Insights and Ideas

  • Relays are SemCom Force Multipliers: Relay-assisted schemes consistently outperform direct transmission, particularly in low Signal-to-Noise Ratio (SNR) regimes. The "diversity gain" from having two paths for the semantic information is significant.
  • Efficiency of Latent Processing: The Learn-and-Forward (LF) strategy is highly efficient. It uses only ~199K parameters compared to the ~26M required for Decode-and-Forward, yet it achieves comparable or better performance.
  • The Vulnerability of "Meaning": Because SemCom relies on learned latent representations, it is uniquely susceptible to structured attacks. Unlike random noise, the Byzantine relay's perturbations are "smart"—they are specifically designed to disrupt the destination's decision-making process.
  • Single-Pass Execution: The proposed attack doesn't require ground-truth labels or slow, iterative optimization at test time. Once trained, the relay can generate devastating perturbations in a single forward pass of its neural network.

Key Results

  • Performance Gains: At low SNR, relay-assisted schemes achieved an absolute improvement of up to 0.4248 in semantic accuracy over direct transmission. They also saw gains of 2.58 dB in PSNR (image quality).
Fig. 2: Semantic accuracy in semantic relay communications.
  • Attack Impact: The Byzantine attack proved highly effective across all SNR levels. Depending on the power constraints placed on the perturbation, the attack reduced semantic accuracy by up to 89.42%.
Fig. 4: Reduction in semantic accuracy due to relay attack.

Limitations and Future Work

The primary limitation acknowledged is the "insider" nature of the threat: the attack assumes the relay has knowledge of the destination's model architecture and parameters (a white-box or gray-box scenario). While realistic for a compromised node in a jointly deployed system, it represents a worst-case scenario.

Future work suggested by the authors includes:

  • Developing robust SemCom designs and defenses specifically tailored to counter insider threats.
  • Exploring more complex network topologies beyond the three-node model.
  • Investigating "black-box" attacks where the relay has less information about the destination's internal AI models.

Summary updated

Abstract

Semantic communication (SemCom) using a relay node is considered, where the relay assists in conveying task-relevant information from a source to a destination, with meaning defined by classification labels. A joint end-to-end design of the source encoder, relay transformation, and destination decoders is developed to optimize both semantic inference accuracy and perceptual reconstruction quality for image data. Relay-assisted schemes are shown to outperform direct transmission, with gains that become more pronounced in low signal-to-noise ratio (SNR) regimes across different relaying strategies. Alongside these performance benefits, however, a key security vulnerability introduced by the relay is identified, and a Byzantine attack is formulated in which the relay injects structured, learned perturbations into its transmitted signal. The attack operates on intermediate communication representations subject to channel effects, with perturbations generated through a learned mapping applied at test time. It is shown to significantly degrade semantic inference accuracy at the destination. These results reveal a clear trade-off between SemCom gains and security risks in relay-assisted systems, motivating the need for secure relay designs that account for insider threats.
BibTeX
@misc{sagduyu2026ai,
  author = {Yalin E. Sagduyu and Tugba Erpek and Aylin Yener},
  title = {{AI-Native Task-Oriented Semantic Communications over a Relay: Benefits and Vulnerabilities}},
  year = {2026},
  eprint = {2611.08155},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.08155},
}
Figure