Deepfake Image and Video Detection Using CNN

The rapid proliferation of synthetic media technology has elevated deepfake detection from an academic research problem to an operational societal challenge. This paper examines the effectiveness of Convolutional Neural Networks (CNNs) in distinguishing authentic facial content from algorithmically manipulated equivalents, across both static images and video sequences. Three CNN architectures — XceptionNet, EfficientNet-B4, and a temporal CNN-LSTM hybrid — are evaluated against three established benchmarks: FaceForensics++ (FF++), CelebDF, and the Deepfake Detection Challenge Dataset (DFDC). This paper addresses the growing threat of synthetic media and deepfake technology, which has evolved from a mere academic problem into a serious societal challenge. The research examines the effectiveness of Convolutional Neural Networks (CNNs) in distinguishing authentic facial content from algorithmically manipulated equivalents, across both static images and video sequences. Three major CNN architectures were evaluated — XceptionNet, EfficientNet-B4, and a temporal CNN-LSTM hybrid — tested against three established benchmarks: FaceForensics++ (FF++), CelebDF, and the Deepfake Detection Challenge Dataset (DFDC). When controlled within-distribution experiments were conducted, meaning the models were trained and tested on the same dataset, the results appeared quite promising —accuracy exceeded 97% in almost all configurations. However, the real test came during cross-dataset evaluation, where the results exposed a serious and systematic performance collapse. XceptionNet, for instance, achieved 99.26% accuracy on FF++, but when tested on DFDC, it dropped drastically to just 51.31%. Even more concerning was the fake video recall, which degraded to 13.16% — barely above random chance. Essentially, the model was close to just guessing.This paper investigates the effectiveness of Convolutional Neural Networks in detecting deepfakes across both images and videos. Three architectures — XceptionNet, EfficientNet-B4, and a CNN-LSTM hybrid — were evaluated on three datasets: FaceForensics++, CelebDF, and the Deepfake Detection Challenge Dataset.Within-distribution testing yielded consistently high accuracy, surpassing 97% across most configurations. However, cross-dataset evaluation revealed a severe performance collapse. XceptionNet dropped from 99.26% accuracy on FF++ to just 51.31% on DFDC, with fake video recall falling to 13.16% — barely above random guessing. This drastic degradation demonstrates that models are not learning genuine manipulation cues but are instead memorising dataset-specific artefacts, making them unable to generalise to unseen data.The paper identifies this generalisation failure as the central unsolved problem in the deepfake detection field. No matter how well a model performs in controlled settings, it struggles significantly when exposed to real-world, out-of-distribution content. The research also highlights the role of video compression in degrading detection performance and raises important ethical concerns around deployment. FFT frequency-domain analysis is proposed as a promising complementary technique to improve robustness. Overall, the study concludes that current models require more diverse training data and stronger generalisation strategies before being considered deployment-ready.The core argument of this paper is that this performance collapse occurs because models trained on narrow datasets learn to recognise dataset-specific artefacts rather than actual manipulation cues. This is a fundamental generalisation problem and remains the biggest unsolved challenge in the deepfake detection field. Additionally, the paper discusses the impact of compression effects on detection performance, the ethical implications of deploying such systems in real-world scenarios, and FFT frequency-domain detection as a promising complementary approach that could improve generalisation. The overall conclusion is that current deepfake detection models are not yet fully ready for real-world deployment and require more diverse training data along with better generalisation-focused strategies to effectively combat the ever-growing threat of synthetic media. Controlled within-distribution experiments produced consistently high accuracy, exceeding 97% in most configurations. Cross-dataset evaluation, however, exposed a systematic and severe performance collapse: XceptionNet fell from 99.26% accuracy on FF++ to 51.31% on DFDC, with fake-video recall degrading to 13.16% — barely above random chance. This paper argues that such collapse reflects the central unsolved problem of the field: models trained on narrow datasets learn to recognise dataset-specific artefacts rather than generalisation-robust manipulation cues. Compression effects, ethical implications of deployment, and frequency-domain detection as a promising complementary approach are additionally discussed.