IoT-Oriented Security for Small Sensor Systems Using DnCNN Denoising and Multimodal Feature Fusion for Image Forgery...

Area of Science:

• Digital Forensics
• Computer Vision
• Artificial Intelligence

Background:

• Increasing use of CCTV, miniature sensors, and IoT devices raises concerns about image authenticity.
• Advanced editing tools and generative models create sophisticated forgeries undetectable by traditional methods.
• Existing algorithms often rely on single forensic cues, limiting their robustness against diverse manipulations.

Purpose of the Study:

• To develop a novel forgery detection framework, MultiFusion, for authenticating sensor-generated images.
• To address the limitations of single-cue forensic methods by integrating complementary image characteristics.
• To enhance the interpretability of forgery detection results.

Main Methods:

• MultiFusion framework integrates SRM-based noise residuals, EfficientNet-B0 hierarchical texture features, and vision transformer global structural relationships.
• A DnCNN denoising layer preprocesses images to suppress sensor noise and preserve tampering traces.
• Explainable AI techniques, combining Grad-cam and transformer attention maps, generate interpretable heatmaps highlighting manipulated regions.

Main Results:

• MultiFusion achieved high detection accuracy of 96.69% on the CASIA 2.0 dataset.
• The framework demonstrated good generalization capabilities across different image types and manipulations.
• Interpretable heatmaps effectively identified regions of image manipulation.

Conclusions:

• The MultiFusion framework offers a robust and explainable solution for image forgery detection.
• Its multimodal feature fusion and normalized denoising advance the authentication of CCTV, sensor, and IoT imagery.
• The integration of explainable AI provides crucial insights into detected manipulations.