The paper addresses the challenge of federated unlearning in Human Activity Recognition (HAR) to protect user privacy. Federated Learning (FL) is widely used to build global HAR models without transmitting raw data, but it faces challenges due to regulations like the General Data Protection Regulation (GDPR), which allow users to request data removal. The authors propose a lightweight machine unlearning method that uses a third-party dataset unrelated to model training to fine-tune the global model by aligning the predicted probability distribution on forgotten data with the third-party dataset. They introduce a membership inference evaluation method to assess the effectiveness of unlearning. Experimental results on diverse datasets show that their method achieves comparable accuracy to retraining methods, with speedups ranging from hundreds to thousands. The main contributions include a lightweight unlearning method and a membership inference evaluation method, which significantly conserve computational resources compared to retraining. The paper also discusses the selection of third-party data and its impact on unlearning effectiveness.The paper addresses the challenge of federated unlearning in Human Activity Recognition (HAR) to protect user privacy. Federated Learning (FL) is widely used to build global HAR models without transmitting raw data, but it faces challenges due to regulations like the General Data Protection Regulation (GDPR), which allow users to request data removal. The authors propose a lightweight machine unlearning method that uses a third-party dataset unrelated to model training to fine-tune the global model by aligning the predicted probability distribution on forgotten data with the third-party dataset. They introduce a membership inference evaluation method to assess the effectiveness of unlearning. Experimental results on diverse datasets show that their method achieves comparable accuracy to retraining methods, with speedups ranging from hundreds to thousands. The main contributions include a lightweight unlearning method and a membership inference evaluation method, which significantly conserve computational resources compared to retraining. The paper also discusses the selection of third-party data and its impact on unlearning effectiveness.