A survey of secure semantic communications

Rui Meng, Song Gao, Dayu Fan, Haixiao Gao, Yining Wang, Xiaodong Xu, Bizhu Wang, Suyu Lv, Zhidi Zhang, Mengying Sun, Shujun Han, Chen Dong, Xiaofeng Tao, Ping Zhang

Introduction

As traditional communication systems approach the theoretical limits of the "Shannon Trap," focusing solely on the accurate reconstruction of raw bits, the transition toward 6G necessitates a paradigm shift. Semantic Communication (SemCom) represents this evolution, moving beyond symbol accuracy to the transmission of "meaning." By utilizing artificial intelligence to extract and transmit only essential semantic features, SemCom promises a revolutionary "Internet of Intelligence." However, this reliance on Deep Learning (DL) introduces a new frontier of security risks. This paper provides a comprehensive roadmap of the vulnerabilities inherent in the SemCom lifecycle and surveys the defensive technologies—from blockchain to physical-layer security—essential for building a trustworthy network.

The Vulnerability of Intelligence

Unlike traditional bitstreams, semantic features are highly structured and model-dependent. This inherent structure makes them susceptible to "semantic-level" attacks that do not exist in classical systems. The survey identifies three primary categories of threat:

  1. Knowledge Base Poisoning: Attackers inject malicious data into the AI’s training set, effectively corrupting the "brain" of the communication system.
  2. Gradient Leakage: During collaborative training (such as Federated Learning), private data can be stolen by analyzing the gradients shared between nodes.
  3. Semantic Jamming: Rather than simply creating noise on a frequency, attackers can interfere with the specific meaning of a message, leading to misinterpretation by the receiver.

Furthermore, the paper highlights the risk of Model Inversion. If an attacker understands the underlying semantic model, they can reconstruct sensitive original data from seemingly abstract transmitted features. This is particularly dangerous in Model Slicing, where models are divided into pieces for edge distribution; the survey notes that specific slices, such as the final layer of a classification model, are high-value targets for label-inference attacks.

The SemCom Security Lifecycle

The paper categorizes defensive research into a structured framework based on the three stages of the SemCom lifecycle:

1. Model Training

Securing the initial creation of semantic models often involves Federated Learning (FL). To protect this stage, researchers employ Data Cleaning and Robust Learning techniques. For instance, geometric median-based aggregation in FL has been shown to maintain high performance even when a significant portion of participating clients are malicious.

2. Model Transfer

As model slices move from the cloud to edge devices, they must be protected against tampering and theft. The survey evaluates the use of Blockchain for integrity and Model Compression to reduce the attack surface. Pruning techniques can achieve compression rates of up to 98% without sacrificing performance, which is vital for securing resource-constrained IoT devices.

3. Information Transmission

Ensuring the privacy of semantic features over wireless channels requires a mix of Cryptography and Physical-Layer Security (PLS). The paper discusses the use of Differential Privacy (DP), which adds noise to protect data. However, it warns of a critical Privacy-Accuracy Trade-off: in SemCom, excessive noise can destroy the "meaning" of the message, leading to total communication failure.

Quantitative Performance and Defensive Success

The survey synthesizes several quantitative benchmarks that demonstrate the feasibility of secure SemCom:

  • Efficiency in Encryption: Research into BatchCrypt (a form of Homomorphic Encryption) shows training speedups of 23–93 times compared to traditional methods while maintaining model accuracy.
  • High-Fidelity Privacy: Fully Homomorphic Encryption (FHE) implementations on ResNet-20 models achieved 92.43% accuracy on encrypted data, nearly matching the 91.89% accuracy of plaintext models.
  • Semantic Robustness: The shift from correcting syntactic noise to managing "semantic noise" (ambiguity) requires "Semantic Soft Matching," a more flexible approach than traditional error correction.

Limitations and Future Directions

Despite recent progress, several hurdles remain for the deployment of secure SemCom in 6G:

  • Computational Overhead: Advanced cryptography like Homomorphic Encryption remains too slow for real-time applications, sometimes requiring seconds for a single multiplication.
  • Explainability: There is a lack of "Explainable Robust Learning." Understanding why a model is vulnerable to specific semantic tweaks is necessary to build better defenses.
  • Dynamic Environments: Current security models are largely static. Future systems require "Intelligent Data Cleaning" that can adapt to shifting semantic contexts in real-time, such as an autonomous vehicle encountering new road signs.
  • Multimodal Security: Securing systems that simultaneously process text, audio, and video remains an open and complex challenge.

Conclusion

Semantic Communication is a cornerstone of the next generation of wireless networks, but its intelligence is a double-edged sword. By systematizing the threats across the model lifecycle and evaluating defensive strategies, this survey provides a foundation for developing the robust, private, and efficient protocols required for the future Internet of Intelligence.

Summary updated

BibTeX
@article{meng2025survey,
  author = {Rui Meng and Song Gao and Dayu Fan and Haixiao Gao and Yining Wang and Xiaodong Xu and Bizhu Wang and Suyu Lv and Zhidi Zhang and Mengying Sun and Shujun Han and Chen Dong and Xiaofeng Tao and Ping Zhang},
  title = {{A survey of secure semantic communications}},
  journal = {J. Netw. Comput. Appl.},
  year = {2025},
  doi = {10.1016/j.jnca.2025.104181},
}
Figure