Deepfakes have posed a major threat to society as the ability to create manipulated and defraud videos which are highly realistic embarking a question mark on the integrity of digital media and trustworthiness of online content and hence makes deepfake detection a necessity. The challenges presented in eye gaze tracking to detect deepfakes have been made up considering the cases focusing on poses and illumination with better accuracy for eye gaze tracking. In this paper, we proposed a model that implements Eye gaze tracking to detect deepfakes using the Improved gazeNet model utilizing the pre-trained ResNet-50 and the standard gazeNet model. The model enhancement includes dual integration, i.e., the integration of weights from ResNet-50 and gazeNet followed by feature integration. Along with eye gaze features, the size and color features of the whole face are extracted. The proposed method is evaluated on the World Leader dataset (WLDR). For deepfake detection, a hybrid classifier combining Long Short Term Memory (LSTM) and Convolution Neural Network (CNN) is trained with the extracted features. The experimental results and findings demonstrate that our proposed model outperforms various approaches such as Deep Belief Networks (DBN), LSTM, Random Forest (RF), Recurrent Neural Networks (RNN), Bidirectional Gated Recurrent unit (Bi-GRU), Support Vector Machine (SVM) to evaluate our deepfake dataset with several performance measures with an accuracy of 91.4 % for database images, 87.3% for specific poses, and 91.8% for illumination conditions.