Video Tampering Detection Based on Deep Learning Techniques

Abstract

Dashcams have become standard in modern vehicles, capturing continuous video footage that can serve as vital digital evidence in legal disputes or insurance claims. However, the reliability of these recordings is threatened by various forms of video tampering, including video forgery, frame deletion, frame insertion, and even deepfake-based alterations. These challenges demand robust and intelligent forgery detection systems capable of identifying manipulations at both frame and video levels. This master thesis presents a deep learning-based approach for automated video tampering detection in dashcam recordings. It begins with an in-depth theoretical analysis of digital video forgery, outlining its common forms and the limitations of

traditional forensic methods. The study then focuses on the use of Convolutional Neu- ral Networks (CNNs), known for their effectiveness in visual anomaly detection,

and proposes a novel detection pipeline built upon the ResNet50 architecture. The main contribution is the implementation of a custom model named BinaryCNN, a CNN-based architecture derived from ResNet50 and fine-tuned to classify dashcam

videos into “original” or “forged” (specifically for frame deletion forgery). The de- tection pipeline includes video preprocessing, frame extraction, individual frame

classification, and score aggregation to determine video-level authenticity. The system was trained and evaluated using the publicly available Video Forgery Dataset (Kaggle), which contains thousands of dashcam videos, divided into classes representing different tampering techniques: frame deletion, insertion, duplication, flipping, rotation, and zooming. For this work, we focused on a binary classification task: detecting frame deletion forgery vs. original. The dataset includes 1181 videos per class, split evenly between training and testing sets, and covers diverse driving conditions. The proposed BinaryCNN model achieved an overall accuracy of 79%, with a precision of 81%, recall of 78%, and F1-score of 79%, demonstrating its effectiveness in identifying forged videos under real-world conditions. These results confirm the capability of CNN-based models to tackle video forgery detection with high reliability. Additionally, a user-friendly web interface was developed using Flask to demonstrate the real-time application of this deep learning video tampering detection system. Users can upload a dashcam video and instantly receive a classification result. This research contributes to the growing field of automated forgery detection

and highlights the potential of deep learning techniques, especially ResNet50-based

CNNs, in building scalable, intelligent, and robust systems for video tampering de- tection.