<p>Federated learning is a new distributed learning framework that can train multi-party user data without knowing the original user data, which protects user privacy to a certain extent. However, the application of federated learning in the Internet of Things still faces some problems, such as the high cost of client-server communication, frequent communication failures, and lack of high-quality data to guarantee the training process. To solve the problems of communication costs affected by different encryption schemes, dynamic user joining and exiting, and low user data quality affecting accuracy, this paper proposes a data security and privacy protection method based on EC-ElGamal in federated learning. EC-ElGamal is an additive homomorphic encryption scheme based on ElGamal elliptic curves, which has a shorter key length and less traffic than the ElGamal scheme. Firstly, the scheme uses lightweight EC-ElGamal and improved multi-key EC-ElGamal cryptosystem in the Internet of Things environment to realize cooperative training of multi-key users under federated learning and reduce communication costs. Secondly, FSOA(Federated sum optimization algorithm) is proposed to obtain high-quality data for training, which can avoid the problem of low data quality affecting accuracy. At the same time, a dynamic user joining and exiting algorithm is proposed, which can realize the dynamic user update without affecting the training accuracy. The experiment results shows that by selecting high-quality training data and a short key length encryption scheme for model training, The total encryption and decryption time of this scheme is lower than the other three encryption and decryption time of Paillier, ElGamal and LWE, the communication cost of the scheme is reduced, and the training accuracy is ensured under the condition of dynamic updating of users.</p>