Recently, data mining over transactional data streams has become an attractive research area. However, releasing raw transactional data streams, in which only explicit identifying information must be removed, may compromise individual privacy. Many privacy-preserving approaches have been proposed for publishing static transactional data. Due to the characteristics of data streams, which must be processed quickly, static data anonymization methods cannot be directly applied to data streams. In this paper, we first analyze the privacy problem in publishing transactional data streams based on a sliding window. Then, we present two dynamic algorithms with generalization and suppression to anonymize continuously a sliding window to make it satisfy ρ-uncertainty by structuring an affected sensitive rules trie, because the removal and addition of transactions may make the current sliding window fail to satisfy ρ-uncertainty. Experimental results show that our methods are more efficient than sliding window anonymization with batch processing by using existing static anonymization methods.

User data usually exists in the organization or own local equipment in the form of data island. It is difficult to collect these data to train better machine learning models because of the General Data Protection Regulation (GDPR) and other laws. The emergence of federated learning enables users to jointly train machine learning models without exposing the original data. Due to the fast training speed and high accuracy of random forest, it has been applied to federated learning among several data institutions. However, for human activity recognition task scenarios, the unified model cannot provide users with personalized services. In this paper, we propose a privacy-protected federated personalized random forest framework, which considers to solve the personalized application of federated random forest in the activity recognition task. According to the characteristics of the activity recognition data, the locality sensitive hashing is used to calculate the similarity of users. Users only train with similar users instead of all users and the model is incrementally selected using the characteristics of ensemble learning, so as to train the model in a personalized way. At the same time, user privacy is protected through differential privacy during the training stage. We conduct experiments on commonly used human activity recognition datasets to analyze the effectiveness of our model.