Abstract: The issue of data island has always been a difficult problem during the development of artificial intelligence. The risk of privacy disclosure in model training and application further impedes the development of artificial intelligence technology. Federated learning, emerging as an efficient means of privacy protection, is a distributed machine learning technique, which enables to train a lossless learning model through local training and parameter transfer of participants without directly obtaining data sources. However, study results show that there are still many security risks in federated learning. Aiming at the security problems in federated learning, this paper analyzes three main security threats, including poisoning attacks, adversarial attacks and privacy disclosure, and summarizes the latest defense measures. Finally, this paper discusses the security issues still existing in the current federated learning with related solutions.