Abstract
Semantic communication extracts useful semantic information from raw data by using a shared semantic knowledge base between transmitters and receivers. This approach reduces data volume and improves transmission efficiency. In multi-user semantic communications, the semantic knowledge base relies on data from multiple users to ensure accurate semantic extraction, a process facilitated by federated learning, which enables collaboration without sharing raw data. However, outdated, irrelevant, or harmful data may compromise the reliability of the semantic knowledge base. To address this challenge, this study proposes an efficient and trusted federated unlearning mechanism for multi-user semantic knowledge bases. The mechanism considers single-layer and multi-layer multi-user semantic communications and introduces a federated unlearning method based on gradient ascent and trusted execution environments; this method effectively eliminates the impact of invalid data while preserving reliability. Experimental results demonstrate that the proposed mechanism significantly reduces computational overhead while maintaining comparable accuracy in unlearned data and the post-unlearning semantic knowledge base.
Introduction
In the era of 6G, Semantic Communication (SemCom) has emerged as a paradigm shift, focusing on transmitting the "meaning" of data rather than raw bits. This is achieved through a shared Semantic Knowledge Base (SKB) between transmitters and receivers. Typically, these SKBs are trained using Federated Learning (FL), allowing multiple users to contribute to a global model without compromising their private raw data.
However, the integrity of the SKB can be threatened by "poisonous" data, which may be outdated, irrelevant, or intentionally malicious. Furthermore, privacy regulations like the "right to be forgotten" necessitate a way to remove a specific user's data influence from the shared model. Traditional FL lacks an efficient deletion mechanism, often requiring a full retraining of the global model from scratch—a process that is prohibitively expensive in terms of time and energy.
System Architecture and Unlearning Mechanism
The proposed framework addresses these challenges by integrating federated unlearning with hardware-based security.
The system architecture encompasses four main stages:
- Initialization and Sharing: The global SKB is initialized and shared among users.
- Semantic Extraction and Coding: Senders use semantic encoders and source-channel coding to transmit information.
- Decoding and Reconstruction: Receivers decode the symbols and reconstruct the semantic information using the shared SKB.
- Efficient Update for Invalid Data: This is the core contribution, where the system identifies and removes the influence of outdated, irrelevant, or malicious data.
Methodology
The unlearning process employs a multi-faceted strategy:
- Gradient Ascent (GA): Instead of the standard gradient descent used for learning, the system applies gradient ascent to the data targeted for removal. This "pushes" the model parameters away from the patterns associated with the unwanted data.
- Compensation Training: To maintain the overall accuracy of the SKB, a small group of "good" users performs a brief round of training to stabilize the model after the GA step.
- Hierarchical Unlearning: Recognizing that real-world networks are multi-layered (Edge -> Region -> Central Cloud), the mechanism is designed to backtrack and correct historical gradients across these layers.
- Trusted Execution Environments (TEE): To ensure the unlearning process is not manipulated by malicious actors, the computation is performed within a TEE. This provides hardware-level assurance that the unlearning algorithm was executed as intended.
Experimental Evaluation
The framework was tested using three standard benchmarks: MNIST, Fashion-MNIST, and CIFAR-10, under both Independent and Identically Distributed (IID) and Non-IID conditions.
Key Results
- Efficiency: The proposed method achieved a significant speedup over full retraining. On the MNIST dataset, unlearning took approximately 57 seconds, compared to 357 seconds for full retraining—a nearly 6x improvement.
- Accuracy Maintenance: Post-unlearning accuracy remained nearly identical to models trained from scratch without the invalid data, proving that the "forgetting" process does not degrade general performance.
- Security and Robustness: The integration of TEEs ensured stability even when 10-20% of users were malicious, outperforming existing methods like FedEraser or Shard which saw significant performance degradation.
- Reduced Communication Overhead: By involving only a subset of users in the compensation phase, the bandwidth required for unlearning was drastically reduced.
Conclusion
This research provides a robust solution for maintaining the integrity and privacy of multi-user semantic knowledge bases. By combining gradient ascent with trusted hardware, the proposed federated unlearning mechanism offers an efficient, secure, and scalable way to manage data in 6G semantic communication networks. Future work will focus on handling device heterogeneity and extending the framework to more complex, large-scale semantic models.
Summary updated