Recent Advances in Deep Learning-Based Source Camera Identification and Device Linking.

Area of Science:

• Digital Forensics
• Computer Vision
• Machine Learning

Background:

• Photo-response non-uniformity (PRNU) is a established technique for camera source identification and device linking in forensic science.
• Recent advancements in deep learning (DL) offer new methods to analyze sensor artifacts for forensic purposes.

Purpose of the Study:

• To review and compare the effectiveness of various deep learning architectures for PRNU-based camera identification.
• To evaluate the performance of DL techniques at both the model and device levels over time.
• To identify challenges in current DL approaches for forensic image analysis.

Main Methods:

• Review of deep learning architectures including CNNs, residual learning, encoder-decoder, dual-branch, and contrastive learning.
• Comparative analysis of DL model performance on camera source identification and device linking tasks.
• Evaluation of effectiveness at model-level and device-level identification.

Main Results:

• Deep learning approaches demonstrate high accuracy at the model level for PRNU analysis.
• Robust device-level identification remains a significant challenge, especially with advanced imaging pipelines.
• Effectiveness is impacted by camera-integrated and AI-driven image enhancements.

Conclusions:

• While DL excels at model-level identification, achieving reliable device-level identification requires further research.
• Evolving photographic capture practices necessitate the development of advanced techniques and updated datasets.
• Future work should focus on overcoming limitations posed by AI-driven enhancements in modern cameras.