CNN-Based Multi-Modal Camera Model Identification on Video Sequences.

Area of Science:

• Multimedia Forensics
• Computer Vision
• Digital Signal Processing

Background:

• Source camera identification is crucial for copyright protection and criminal investigations.
• Existing methods often rely on single data modalities (visual or audio), limiting accuracy.
• The need for robust camera model identification in diverse video sources is increasing.

Purpose of the Study:

• To develop and evaluate novel multi-modal methods for camera model identification in video sequences.
• To compare the performance of multi-modal approaches against traditional mono-modal techniques.
• To assess the effectiveness of these methods on both native and social media uploaded videos.

Main Methods:

• Development of two Convolutional Neural Network (CNN)-based camera model identification methods.
• Implementation of a novel multi-modal scenario integrating both audio and visual information.
• Testing on the Vision dataset, comprising nearly 2000 video sequences from various devices.

Main Results:

• The proposed multi-modal methods demonstrated significantly superior performance compared to mono-modal counterparts.
• Effectiveness was validated across native videos and those shared on platforms like YouTube and WhatsApp.
• The joint exploitation of audio and visual data proved to be a valuable strategy for camera identification.

Conclusions:

• Multi-modal approaches offer a substantial advancement in video camera model identification.
• These methods provide a robust solution for multimedia forensic applications.
• Future research can explore more complex scenarios and data fusion techniques.