As a novel distributed machine learning scheme, federated learning (FL) efficiently realizes the collaborative training of models by global participants while also protecting their data privacy. Due to the independence of participants' local data and the inability of the FL server to access the local data, many IIoT applications with strong data sensitivity are increasingly incorporating FL technology. However, it also exposes a great security vulnerability. Malicious adversaries manipulate local data to perform covert targeted poisoning attacks or other harmful behaviors to affect the global model. In addition, due to the diversity of IIoT data in actual scenarios, different data distribution scenarios can also cause different attack effects In this work, we devise a defense technique called FDSFL against multiple malicious adversaries and various targeted poisoning attacks involving both IID and non-IID data distribution scenarios. It runs on the server-side and mainly includes three execution modules: pairwise cosine similarity, clustering mechanism, and filtering strategy, which can dynamically filter malicious updates during the iterative training process. We demonstrate that our designed FDSFL outperforms the state-of-the-art in maintaining global model accuracy and reducing attack success rates through extensive experiments on three general datasets.