Deepfake technology, a burgeoning concern across diverse sectors, necessitates robust detection methodologies to counter its potential threats. This research delves into the realm of deepfake detection, exploring various approaches including custom convolutional neural network (CNN) architectures, transfer learning with pre-trained models such as Xception, and a hybrid model amalgamating Xception with Long Short-Term Memory (LSTM). Additionally, an explainable AI component is introduced to provide transparency and insight into model decisions. The efficacy of these methodologies is evaluated using a dataset comprising faces extracted from the deepfake-detection-challenge, with results showcasing the superiority of the hybrid LSTM-Xception model, achieving a higher accuracy on the test set. The integration of explainable AI techniques enhances trust and interpretability, contributing to the ongoing efforts to develop robust deepfake detection mechanisms.