SCU: An Efficient Machine Unlearning Scheme for Deep Learning Enabled Semantic Communications

Weiqi Wang, Zhiyi Tian, Chenhan Zhang, Shui Yu

Introduction

Semantic Communication (SC) represents a paradigm shift in networking, focusing on transmitting the underlying meaning of data rather than raw bits. This is achieved through Deep Learning (DL) models, specifically joint encoder-decoder pairs (codecs). However, as these systems handle sensitive personal data—such as biometric information from edge device sensors—compliance with privacy regulations like GDPR becomes essential. Specifically, the "right to be forgotten" requires efficient machine unlearning techniques to erase specific user data from trained models without the prohibitive cost of full retraining.

The Challenge of Unlearning in Semantic Communications

Standard machine unlearning techniques, typically designed for supervised classification, face significant hurdles in SC systems:

  1. Joint Dependency: The tight coupling between encoders and decoders means modifying one can easily degrade the performance of the entire communication link.
  2. Unsupervised Nature: SC models often utilize unsupervised methods like Variational Autoencoders (VAEs), making label-based unlearning inapplicable.
  3. Quantification: Measuring "forgetting" in reconstruction-based codecs is inherently difficult compared to classification tasks.

The SCU Framework

To address these challenges, the authors propose SCU (Semantic Communication Unlearning). The framework operates in two primary phases to ensure efficient data erasure while maintaining system utility.

The SCU architecture featuring Joint Unlearning and Contrastive Compensation.
  1. Joint Unlearning (JU): Instead of a full retrain, SCU updates the encoder and decoder simultaneously. It minimizes the Mutual Information (MI) between the learned semantic representations ($z_e$) and the samples to be erased ($x_e$). To prevent catastrophic forgetting of remaining data, a Kullback-Leibler Divergence (KLD) constraint keeps the unlearned model weights close to the original "knowledgeable" model.
  2. Contrastive Compensation (CC): To recover performance lost during unlearning, a contrastive loss is employed. Erased data are treated as negative samples, while remaining data are positive samples. This forces the model to distinguish between what must be forgotten and what must be retained, reinforcing "clean" knowledge.

Experimental Evaluation

The researchers evaluated SCU using MNIST, CIFAR10, and CIFAR100 datasets. Effectiveness was measured using a backdoor-based evaluation: data to be forgotten was "poisoned" with a visual patch. Successful unlearning is indicated by a downstream classifier's inability to detect this backdoor in reconstructed images, even if the visual quality remains high.

Performance comparison of SCU against baselines (VBU, HBU) across MNIST, CIFAR10, and CIFAR100.

Key Results

  • High Effectiveness: In tests on MNIST, the "backdoor accuracy" (a proxy for how much the model remembers the erased data) dropped from 97.72% to roughly 7%, indicating successful erasure.
  • Utility Preservation: While baseline unlearning methods (like HBU or VBU) often caused the model's accuracy on clean data to crash (sometimes to near 0%), SCU maintained clean data accuracy at 97.56% on MNIST.
  • Efficiency: SCU is significantly faster than retraining from scratch. It is over 50x faster than Hessian-based unlearning (HBU) on complex datasets like CIFAR100, proving its viability for edge-device deployment.

Conclusion

SCU provides a robust and efficient solution for the "right to be forgotten" in semantic communication systems. By leveraging Mutual Information for precision erasure and contrastive learning for stability, it maintains the critical bond between encoder and decoder. Future work will explore task-oriented unlearning (e.g., for speech or VQA) and dynamic, streaming unlearning requests in live network environments.

Summary updated

Abstract

Deep learning (DL) enabled semantic communications leverage DL to train encoders and decoders (codecs) to extract and recover semantic information. However, most semantic training datasets contain personal private information. Such concerns call for enormous requirements for specified data erasure from semantic codecs when previous users hope to move their data from the semantic system. Existing machine unlearning solutions remove data contribution from trained models, yet usually in supervised sole model scenarios. These methods are infeasible in semantic communications that often need to jointly train unsupervised encoders and decoders. In this paper, we investigate the unlearning problem in DL-enabled semantic communications and propose a semantic communication unlearning (SCU) scheme to tackle the problem. SCU includes two key components. Firstly, we customize the joint unlearning method for semantic codecs, including the encoder and decoder, by minimizing mutual information between the learned semantic representation and the erased samples. Secondly, to compensate for semantic model utility degradation caused by unlearning, we propose a contrastive compensation method, which considers the erased data as the negative samples and the remaining data as the positive samples to retrain the unlearned semantic models contrastively. Theoretical analysis and extensive experimental results on three representative datasets demonstrate the effectiveness and efficiency of our proposed methods.
BibTeX
@article{wang2025scu,
  author = {Weiqi Wang and Zhiyi Tian and Chenhan Zhang and Shui Yu},
  title = {{SCU: An Efficient Machine Unlearning Scheme for Deep Learning Enabled Semantic Communications}},
  journal = {IEEE Trans. Inf. Forensics Secur.},
  year = {2025},
  doi = {10.1109/tifs.2024.3516576},
}
Figure