Due to the client’s malicious or semi-honest tendencies, the final result of the model may be different when the Federated Learning model is trained. Accurately assessing the contribution of individual clients in federated learning is essential to screen out good clients and weed out bad ones. And since the client’s data is non-id, this can also lead to potential interference. All of the above problems bring challenges to model training, especially because the nature of federated learning means that we cannot touch the client’s local data. In this paper, we propose a Robust, unbiased, and K-means-based client contribution evaluation (FedSKM) method to measure client contributions while filtering out the negative impact of malicious clients. The original Fedruth only considered the boosting attack, and could only cull the impact of models that were far from normal inference, but not the impact of free-riding, and would result in the freerider’s contribution not being lower than that of the normally trained client. Our approach successfully identifies free-riders who only perform weighted aggregation of global and local models from the beginning, and the contribution of the free-rider Boosting Attack client is reduced at the same time, so the client’s contribution can be quantified robustly.