AI system detects manipulated video frames with 95% accuracy

With the rapid spread of digital content, doctored videos pose growing risks across media, security, and legal domains. A new study published in The Journal of Engineering Research introduces an automated approach to detect interpolated frames—artificially inserted images used to smooth manipulated video sequences and make them appear authentic.

The research proposes a three-phase pipeline combining classical image processing techniques with deep learning. In the first phase, videos are decomposed into individual frames, each resized to 256 × 256 pixels and filtered to reduce noise. The second phase applies bilinear interpolation to generate synthetic frames, enabling the system to learn how manipulated content behaves.

In the final phase, a convolutional neural network (CNN) analyzes each frame, automatically extracting visual features and classifying frames as original or manipulated. Unlike traditional methods, CNNs learn directly from pixel data, eliminating the need for manual feature design.

The system achieved an accuracy of 95% when tested on the widely used UCF101 dataset, which contains over 13,000 video clips across 101 categories. In comparison, a support vector machine (SVM) model tested under the same conditions reached only 70% accuracy, highlighting the superiority of deep learning in detecting subtle visual manipulations.

The researchers note that the proposed method not only detects manipulated frames but can also pinpoint their exact position within a video stream, enhancing its value for digital forensics and verification.