With the continuous advancement of IoT technology, IoT devices have been widely used in daily life. Due to the need for IoT devices to collect personal data, which often contains sensitive information such as user health status, users are unwilling to directly upload raw data. Federated learning are often adopted for multiple participants training the same model without leaking local raw data. However, some users may upload randomly generated models, known as poisoning models, in order to save computation. These models can cause serious harm to the model updates of federated learning. Therefore, there is a privacy-preserving step-size collection solution to collect users from IoT devices and remove abnormal models from them. The proposed solution achieves n-source anonymity by proposing an anonymous data aggregation protocol based on XOR homomorphic encryption. The analysis shows that the solution efficiency is significantly improved and meets the needs of users.