An Effective Training Strategy for Enhanced Source Camera Device Identification

Source camera identification is a forensic problem of linking an image in question to the camera used to capture it. This could be a useful tool in forensic applications to identify potential suspects of cyber-crime. Over the last decade, several successful attempts have been made to identify the source camera using deep learning. However, existing techniques that provide effective solutions for camera model identification fail to distinguish between different devices of the same model. This is because cameras of different brands and models were used to train the data-driven system when dealing with exact device identification. We show that training the data-driven system on different camera models opens side-channel information on model-specific features, which acts as interference for identifying individual devices of the same model. Thus, we provide an effective training strategy that involves a way to construct the dataset for enhanced source camera device identification. The experimental results suggest that involving only cameras of the same model for training improves the discriminative capability of the data-driven system by eliminating the threat of interfering model-specific features.