The fast development in deep learning techniques, besides the wide spread of social networks, facilitated fabricating and distributing images and videos without prior knowledge. This paper developed an evolutionary learning algorithm to automatically design a convolutional neural network (CNN) architecture for deepfake detection. Genetic algorithm (GA) based on residual network (ResNet) and densely connected convolutional network (DenseNet) as building block units for feature extraction versus multilayer perceptron (MLP), random forest (RF) and support vector machine (SVM) as classifiers generates different CNN structures. A local search mutation operation proposed to optimize three layers:(batch normalization, activation function, and regularizes). This method has the advantage of working on different datasets without preprocessing. Findings using two datasets evidence the efficiency of the suggested approach where the generated models outperform the state-of-art by increasing 1% in the accuracy; this confirms that intuitive design is the new direction for better generalization.