Federated learning (FL) transforms machine learning by allowing decentralized devices to train a global model while protecting data privacy. However, this decentralized approach creates vulnerabilities for adversarial attacks, particularly model poisoning, in which adversaries inject malicious data to compromise the global model. This paper discusses detecting and mitigating label-flipping attacks using UMAP-based dimensionality reduction and K-Means clustering where the model updates are subjected to UMAP dimensionality reduction and obtain embeddings for each model update, then introduce Gaussian noise and use reputation-based outlier detection to deal with noisy data. Extensive evaluations of clean and noisy test data demonstrate the effectiveness of our framework, which is critical for securing federated learning across sensitive domains. This strategy goes through two layers which ensures the accuracy of detection. Thus, this novel algorithm helps in the identification and clearance of malicious clients. Our findings highlight the importance of mitigating adversarial risks in collaborative machine learning systems, thereby advancing the deployment of FL in real-world scenarios.