The advancement of technology and Artificial Intelligence (AI) has posed challenges to online security. While AI has numerous benefits, it comes with its fair share of shortcomings. Deepfakes refer to manipulated media like video, image, and audio. The existence of the Deepfake technology is a notable example of how AI has the potential to pose risks to society. Deep Learning (DL) models have proven effective in differentiating genuine images from synthesized ones. When there is a limited amount of training data, transfer learning is employed wherein a pre-trained DL model is used. Though DL models are predominantly used, there is a need to eliminate features that are redundant from the feature set. To achieve that, in this work, the functionality and workings of the Particle Swarm Optimization (PSO) algorithm for feature selection have been studied. First, features have been extracted using a pre-trained DL model. Thereafter, optimal features are derived from the extracted features by implementing the PSO algorithm. This feature subset is then used to perform classification. The model has been evaluated on the OpenForensics public dataset while using, on average, 44.18% of the feature set. Classification was performed using an artificial neural network (ANN). The selected features using PSO drastically improved the model’s performance from an accuracy of 63.4% (without PSO) to 79.1%.