Delivery Image Forensics Testbed: A Practical Framework to Combat Generative AI Forgery

The Delivery Image Forensics Testbed is an image forensics experimental framework designed for logistics scenarios. It integrates cutting-edge detection models such as CAT-Net, MVSS-Net, and PSCC-Net to identify and locate generative AI forgery traces in delivery item images, addressing security issues like fraud and voucher tampering caused by AI forgery in logistics. The project is from GitHub, original author is abstract1729, released on May 24, 2026.

With the development of generative AI technologies (e.g., Stable Diffusion, Midjourney), the quality of image forgery has reached a level indistinguishable to the naked eye, posing serious risks in logistics scenarios: malicious users forge package damage images to defraud insurance, or tamper with delivery vouchers to cover up dereliction of duty. Traditional forensics methods target low-level features (such as JPEG compression artifacts) and are ineffective against new content generated by AI; moreover, logistics scenarios require rapid screening of massive images, placing high demands on algorithm accuracy and efficiency.

The core missions of the delivery-forensics-testbed project include:

1. Integrate state-of-the-art image forensics models in the field of computer vision;
2. Optimize for the characteristics of delivery images (specific shooting angles, common packaging, typical damage patterns);
3. Establish standardized test datasets and evaluation metrics to objectively measure the practical performance of models;
4. Precisely locate tampered areas to provide visual clues for manual review.

The project integrates three SOTA models:

• CAT-Net: Uses differences in JPEG compression artifacts, combines RGB pixel domain and DCT frequency domain information to detect tampered areas, enabling recognition of secondary processing traces;
• MVSS-Net: Multi-view and multi-scale supervision, optimizes image-level classification, edge-aware supervision, and pixel-level segmentation simultaneously to improve the accuracy of tamper boundary detection;
• PSCC-Net: Progressive spatial-channel correlation learning, gradually refines tamper localization through HRNet backbone network and progressive non-local correlation module, adapting to complex logistics scenarios.

Technical implementation focuses on engineering deployment:

• Cross-platform compatibility: Supports Apple Silicon (MPS backend) and CUDA/CPU/MPS weight conversion, adapting to the heterogeneous IT environments of logistics enterprises;
• Modular design: Each model has independent interface specifications, facilitating performance comparison, model selection, and integration of new models;
• Standardized inference process: Each model directory contains a detailed README_INFERENCE.md, lowering the threshold for non-technical personnel to use.

Practical application scenarios include:

• Insurance fraud detection: Automatically screen AI forgery traces in claim photos, mark them and transfer to manual review;
• Delivery voucher verification: Real-time analysis of signed photos of high-value items to ensure the authenticity of vouchers;
• Dispute evidence review: Provide objective technical analysis basis for delivery disputes to assist arbitration.

Current limitations:

1. Vulnerability to adversarial attacks: Easily deceived by carefully designed perturbations;
2. Lag in new forgery technologies: Difficult to respond promptly to rapidly iterating AI forgery methods;
3. High computational resource requirements: Some models (e.g., PSCC-Net) are difficult to deploy on edge devices. Future directions:
4. Continuous learning mechanism to adapt to new forgeries;
5. Develop lightweight models to adapt to edge devices;
6. Multi-modal fusion (image + metadata + sensor data);
7. Blockchain-based evidence storage to build an untamperable evidence chain.

The delivery-forensics-testbed project is an important step in the application of image forensics technology from academia to practice. By integrating cutting-edge models and adapting to logistics scenarios, it provides technical support for the industry to address AI forgery challenges. In the arms race between generative AI and detection technology, such scenario-based solutions will become key infrastructure to maintain digital trust. We look forward to more solutions to safeguard the authenticity of the digital world.