Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Privacy-Preserving Semantic Communications via Multi-Task Learning and Adversarial Perturbations

Yalin E. Sagduyu, Tugba Erpek, Aylin Yener, Sennur Ulukus

TL;DR

The paper tackles semantic leakage in learned semantic communications by introducing a unified framework that combines multi-task end-to-end learning with competitive min-max training against an adaptive eavesdropper and an auxiliary cooperative perturbation layer. The transmitter–receiver pair learns to support semantic inference and reconstruction while actively limiting the eavesdropper’s ability to infer task-relevant information, with privacy control governed by a weight $w_P$. Empirical results on MNIST and CIFAR-10 over Rayleigh fading plus AWGN channels show that the min-max objective reduces Eve’s performance without sacrificing Bob’s task accuracy and reconstruction quality; the perturbation layer further tightens leakage, even when the legitimate link is trained solely for its own task. The findings offer deployment-oriented guidelines for tunable, end-to-end privacy in realistic wireless settings, with robust performance across latent dimensions, SNRs, and adversarial perturbation strategies, and point to future work on richer tasks, stronger adversaries, and cross-layer integration.

Abstract

Semantic communications conveys task-relevant meaning rather than focusing solely on message reconstruction, improving bandwidth efficiency and robustness for next-generation wireless systems. However, learned semantic representations can still leak sensitive information to unintended receivers (eavesdroppers). This paper presents a deep learning-based semantic communication framework that jointly supports multiple receiver tasks while explicitly limiting semantic leakage to an eavesdropper. The legitimate link employs a learned encoder at the transmitter, while the receiver trains decoders for semantic inference and data reconstruction. The security problem is formulated via an iterative min-max optimization in which an eavesdropper is trained to improve its semantic inference, while the legitimate transmitter-receiver pair is trained to preserve task performance while reducing the eavesdropper's success. We also introduce an auxiliary layer that superimposes a cooperative, adversarially crafted perturbation on the transmitted waveform to degrade semantic leakage to an eavesdropper. Performance is evaluated over Rayleigh fading channels with additive white Gaussian noise using MNIST and CIFAR-10 datasets. Semantic accuracy and reconstruction quality improve with increasing latent dimension, while the min-max mechanism reduces the eavesdropper's inference performance significantly without degrading the legitimate receiver. The perturbation layer is successful in reducing semantic leakage even when the legitimate link is trained only for its own task. This comprehensive framework motivates semantic communication designs with tunable, end-to-end privacy against adaptive adversaries in realistic wireless settings.

Privacy-Preserving Semantic Communications via Multi-Task Learning and Adversarial Perturbations

TL;DR

The paper tackles semantic leakage in learned semantic communications by introducing a unified framework that combines multi-task end-to-end learning with competitive min-max training against an adaptive eavesdropper and an auxiliary cooperative perturbation layer. The transmitter–receiver pair learns to support semantic inference and reconstruction while actively limiting the eavesdropper’s ability to infer task-relevant information, with privacy control governed by a weight . Empirical results on MNIST and CIFAR-10 over Rayleigh fading plus AWGN channels show that the min-max objective reduces Eve’s performance without sacrificing Bob’s task accuracy and reconstruction quality; the perturbation layer further tightens leakage, even when the legitimate link is trained solely for its own task. The findings offer deployment-oriented guidelines for tunable, end-to-end privacy in realistic wireless settings, with robust performance across latent dimensions, SNRs, and adversarial perturbation strategies, and point to future work on richer tasks, stronger adversaries, and cross-layer integration.

Abstract

Semantic communications conveys task-relevant meaning rather than focusing solely on message reconstruction, improving bandwidth efficiency and robustness for next-generation wireless systems. However, learned semantic representations can still leak sensitive information to unintended receivers (eavesdroppers). This paper presents a deep learning-based semantic communication framework that jointly supports multiple receiver tasks while explicitly limiting semantic leakage to an eavesdropper. The legitimate link employs a learned encoder at the transmitter, while the receiver trains decoders for semantic inference and data reconstruction. The security problem is formulated via an iterative min-max optimization in which an eavesdropper is trained to improve its semantic inference, while the legitimate transmitter-receiver pair is trained to preserve task performance while reducing the eavesdropper's success. We also introduce an auxiliary layer that superimposes a cooperative, adversarially crafted perturbation on the transmitted waveform to degrade semantic leakage to an eavesdropper. Performance is evaluated over Rayleigh fading channels with additive white Gaussian noise using MNIST and CIFAR-10 datasets. Semantic accuracy and reconstruction quality improve with increasing latent dimension, while the min-max mechanism reduces the eavesdropper's inference performance significantly without degrading the legitimate receiver. The perturbation layer is successful in reducing semantic leakage even when the legitimate link is trained only for its own task. This comprehensive framework motivates semantic communication designs with tunable, end-to-end privacy against adaptive adversaries in realistic wireless settings.
Paper Structure (15 sections, 6 figures)

This paper contains 15 sections, 6 figures.

Table of Contents

  1. Introduction
  2. Semantic Communications with Multi-Task Learning
  3. Encoder and Decoder at the Transmitter-Receiver Pair
  4. Multi-task Learning Objective
  5. Semantic Communications Performance
  6. Privacy-Preserving Semantic Communications Against Eavesdropper
  7. Threat Model and Eavesdropper Architecture
  8. Min-max Training Formulation for Privacy Protection
  9. Performance of Privacy Protection with Min-max Training
  10. Adversarial Perturbations as an Auxiliary Mechanism for Privacy
  11. Perturbation Threat Model and Operational Setup
  12. Adversarial Perturbation Mechanisms
  13. Performance of Privacy Protection with Adversarial Perturbations
  14. Future Directions
  15. Conclusion

Figures (6)

  • Figure 1: Privacy-preserving semantic communications: eavesdropping and protection mechanisms.
  • Figure 2: System model for privacy-preserving semantic communications and underlying encoder and decoder architectures.
  • Figure 3: Semantic communications performance.
  • Figure 4: Semantic accuracy during training for different Eve evaluation SNRs (when CIFAR-10 is used and privacy protection weight $w_P$ is 10).
  • Figure 5: Protection against eavesdropping for semantic inference.
  • ...and 1 more figures