The rapid evolution of Internet of Things (IoT) technology has led to the widespread adoption of Human Activity Recognition (HAR) in various daily life domains. Federated Learning (FL) has emerged as a popular approach for building global HAR models by aggregating user contributions without transmitting raw individual data. While FL offers improved user privacy protection compared to traditional methods, challenges still exist.
One particular challenge arises from regulations like the General Data Protection Regulation (GDPR), which grants users the right to request data removal. This poses a new question for FL: How can a HAR client request data removal without compromising the privacy of other clients?
In response to this query, we propose a lightweight machine unlearning method for refining the FL HAR model by selectively removing a portion of a client’s training data. Our method leverages a third-party dataset that is unrelated to model training. By employing KL divergence as a loss function for fine-tuning, we aim to align the predicted probability distribution on forgotten data with the third-party dataset.
Additionally, we introduce a membership inference evaluation method to assess the effectiveness of the unlearning process. This evaluation method allows us to measure the accuracy of unlearning and compare it to traditional retraining methods.
To validate the efficacy of our approach, we conducted experiments using diverse datasets. The results demonstrate that our method achieves unlearning accuracy that is comparable to retraining methods. Moreover, our method offers significant speedups, ranging from hundreds to thousands.
Expert Analysis
This research addresses a critical challenge in federated learning, which is the ability for clients to request data removal while still maintaining the privacy of other clients. With the increasing focus on data privacy and regulations like GDPR, it is crucial to develop techniques that allow individuals to have control over their personal data.
The proposed lightweight machine unlearning method offers a practical solution to this challenge. By selectively removing a portion of a client’s training data, the model can be refined without compromising the privacy of other clients. This approach leverages a third-party dataset, which not only enhances privacy but also provides a benchmark for aligning the predicted probability distribution on forgotten data.
The use of KL divergence as a loss function for fine-tuning is a sound choice. KL divergence measures the difference between two probability distributions, allowing for effective alignment between the forgotten data and the third-party dataset. This ensures that the unlearning process is efficient and accurate.
The introduction of a membership inference evaluation method further strengthens the research. Evaluating the effectiveness of the unlearning process is crucial for ensuring that the model achieves the desired level of privacy while maintaining performance. This evaluation method provides a valuable metric for assessing the accuracy of unlearning and comparing it to retraining methods.
The experimental results presented in the research showcase the success of the proposed method. Achieving unlearning accuracy comparable to retraining methods is a significant accomplishment, as retraining typically requires significant computational resources and time. The speedups offered by the lightweight machine unlearning method have the potential to greatly enhance the efficiency of FL models.
Future Implications
The research presented in this article lays the groundwork for further advancements in federated learning and user privacy protection. The lightweight machine unlearning method opens up possibilities for other domains beyond HAR where clients may need to request data removal while preserving the privacy of others.
Additionally, the use of a third-party dataset for aligning probability distributions could be extended to other privacy-preserving techniques in federated learning. This approach provides a novel way to refine models without compromising sensitive user data.
Future research could explore the application of the proposed method in more complex scenarios and evaluate its performance in real-world settings. This would provide valuable insights into the scalability and robustness of the lightweight machine unlearning method.
In conclusion, the lightweight machine unlearning method proposed in this research offers a promising solution to the challenge of data removal in federated learning. By selectively removing a client’s training data and leveraging a third-party dataset, privacy can be preserved without compromising the overall performance of the model. This research paves the way for further advancements in privacy-preserving techniques and opens up possibilities for the application of federated learning in various domains.