Frame Duplication Detection Using CNN-Based Features with PCA and Agglomerative Clustering

In a court of law, surveillance videos and recordings are the major sources of evidence for any incident or crime. With a simple video editor, the reality could be easily manipulated. This introduces the challenge of verifying the authenticity of contents before they may be used in any critical application domains. In this paper, a video forensic technique is proposed to detect frame duplication forgery in surveillance videos. The proposed technique utilizes convolutional neural network for integrity embedding generation and matching, for video frames duplication detection and localization. Through this work, we compare the efficacy of various deep learning models in generating true embeddings of surveillance footages. To perform experiments a dataset of over 100 authentic and 300 forged, high-resolution HEVC-coded video clips is prepared from surveillance clips. Experimental results indicate that the proposed technique is a good replacement of traditional hand-crafted and compression domain feature-based approaches.