The deepfake has grown faster with generative artificial intelligence which has made it much easier to produce realistic manipulated images and videos. These synthetic media raise serious concerns about the trustworthiness of information, digital security, personal privacy, and overall societal confidence. Traditional methods for detecting deepfakes depend on centralized deep learning systems that need to gather large amounts of images on a single server. To overcome the traditional challenges, this work introduces a new deepfake detection system designed to protect privacy by combining Federated Learning (FL) with Deep Learning (DL) and forensic analysis techniques. The system uses FL, which allows models to train on data stored locally across many devices without transferring the raw images anywhere. It’s built on the ResNet-18 architecture, chosen because it balances good detection capabilities with reasonable computational requirements. ResNet-18 residual learning helps the model train steadily and extract useful features, which fits well with the varied and distributed data environments in federated settings. Experiments are conducted using the FaceForensics++ dataset, a standard resource in media forensics studies which cover four types of manipulations DeepFakes, Face2Face, FaceSwap, and Neural Textures. Since client data distributions differ, which can slow down or destabilize federated training, the FedProx algorithm is used. FedProx adds a term to the local loss function to keep client model updates close to the global model, which helps reduce divergence between clients and improves overall training stability. After local training, only the updated model weights are sent to a central server, where they are averaged to create an improved global model. The experimental results show that the accuracy of 97.20%, precision 96.30%, recall of 96.72, f1-score of 96.90% to keep track of how well the model is learning and generalizing. This study demonstrates a unified framework integrating FL, CNNs, and explainability to enable secure, privacy-preserving, and interpretable deepfake detection, enhancing transparency and user trust.