As the prevalence of advanced face manipulation technologies continues to grow, the detection of deepfakes has emerged as a pivotal area of research, presenting ethical and security challenges. This study introduces an enhanced deepfake detection framework that integrates the EfficientNetB0 model, cutting-edge convolutional neural network (CNN) architecture, with an added long short-term memory (LSTM) layer. The CNN component, EfficientNetB0, excels at extracting spatial features from individual frames, capturing fine-grained details and artifacts indicative of manipulation. The LSTM layer processes sequential dependencies across frames, leveraging temporal inconsistencies that are often present in deepfake videos but difficult to detect from individual images alone. By combining spatial and temporal feature learning, this hybrid CNN-LSTM approach enhances the system’s ability to distinguish real from manipulated media with greater accuracy. Leveraging video frame extraction and comprehensive data augmentation techniques, the system preprocesses inputs to improve generalization on limited training data. The EfficientNetB0 model, pre-trained on the ImageNet dataset, serves as the backbone for feature extraction, utilizing its highly efficient architecture, which includes depth wise separable convolutions. An evaluation on the Celeb-DF dataset demonstrates that the proposed system maintains high accuracy and robustness in detecting deepfake content, while preserving computational efficiency, making it well-suited for real-world applications. The experimental results validate the effectiveness of this extended approach, emphasizing its potential to significantly contribute to the mitigation of the detrimental impacts of deepfakes.