Image Authenticity Detection System Based on CASIA Dataset and CNN

This project is an open-source project for binary classification of image authenticity using Convolutional Neural Networks (CNN), based on the CASIA image tampering detection dataset, demonstrating the application of deep learning in digital image forensics. Developed by Min-Thant-Hein-17 and released on May 24, 2026, the GitHub link is https://github.com/Min-Thant-Hein-17/CASIA_REAL-FAKE_Image_Detector.

With the development of generative AI technology, deepfakes and AI-generated images have become rampant, making image authenticity detection a critical field. This project is a final project for an advanced machine learning course, focusing on solving this problem using deep learning.

The CASIA dataset is a benchmark dataset for digital image forensics, containing real and forged images that cover tampering techniques such as copy-paste, splicing, and deletion. Forged images are processed with JPEG compression, noise addition, etc., to simulate real-world scenarios and improve the model's generalization ability.

The project uses CNN as the core architecture, whose advantages include: local perception to capture local image features, weight sharing to reduce parameters and overfitting risk, and hierarchical extraction of features from low-level to high-level.

Image authenticity detection is modeled as a binary classification problem: class 0 for real images and class 1 for forged images. The output layer uses a Sigmoid activation function to output a probability value between 0 and 1, and authenticity is determined via a threshold (usually 0.5).

Data Preprocessing: Uniform image size, pixel normalization, data augmentation (rotation, flipping, cropping).

Model Training: Uses cross-entropy loss function, Adam optimizer, early stopping to prevent overfitting, and validation set monitoring to select the optimal model.

Deployment: Code is provided in Jupyter Notebook format. No weight files are uploaded; users can run all cells to generate a model.keras model for inference.

This system can be applied in:

1. News Media Review: Quickly screen suspicious content to improve review efficiency;
2. Judicial Forensics: Provide technical support for forensic identification and recognize tampered evidence;
3. Social Media Security: Curb the spread of false information;
4. Finance and Identity Authentication: Serve as a supplementary security layer for liveness detection and face recognition.

Current Limitations:

1. Vulnerability to adversarial attacks: Minor pixel perturbations may lead to misjudgment;
2. Unknown tampering types: Performance degrades when facing new forgery methods;
3. Computational resource requirements: GPU support is needed for training and inference.

Future Directions:

• Multimodal fusion: Combine image metadata and EXIF information;
• Transformer architecture: Explore the application of Vision Transformer;
• Federated learning: Collaborative training to protect privacy;
• Interpretability: Provide visual explanations to enhance credibility.

This project demonstrates a typical application paradigm of deep learning in the field of image authenticity detection, combining the CASIA dataset and CNN architecture to provide a practical technical solution. As generative AI evolves, the challenges of image authenticity detection become more complex. Such open-source projects not only provide code implementations but also serve as a starting point for researchers and developers to understand and improve image forensics technology.