The use of deepfake and synthetic media technologies has come a long way in the past few years, posing serious problems for the authentication of digital media and the prevention of misinformation. This work presents a novel Hybrid CNN–ANN approach to deepfake video detection, combining the advantages of deep spatial feature extraction and effective binary classification. The proposed method will use a sub-set of the DeepFake Detection Challenge (DFDC) dataset to train and test their models. First, input videos are sampled uniformly in time into temporal frames and then they are decoded. Denoising, CLAHE based contrast enhancement and image sharpening techniques are employed to extract and enhance face and head–neck regions. Discriminative 2048 dimensional deep feature representations are extracted from each frame using ResNet50, which is pre-trained on the ImageNet dataset. Multiple dense layers and a sigmoid output unit form a Multi-Layer Perceptron (MLP) based Artificial Neural Network (ANN) implemented for classification of the extracted features. A multi-level class balancing strategy is employed, which is based on stratified video-level splitting, controlled frame-level under-sampling and data augmentation, to overcome the issue of severe class imbalance. The proposed framework is evaluated in a realistic manner on the DFDC test set, which shows the high accuracy of 95.13% and an AUC of 0.954. The result shows that the proposed Hybrid CNN–ANN architecture is able to classify authentic and manipulated videos effectively with a moderate computational cost and high classification accuracy.