• 

  Types of forgeries our model can detect

• 

  Accuracy of the model

• 

  Fake image for testing

• 

  UI of the deployed model

• 

  Output of the model

Image Forgery Detection Using Machine Learning - The AI Hackathon

Description:

In this project, we present a cutting-edge method for Image Forgery Detection using a novel combination of Error Level Analysis (ELA) and Convolutional Neural Networks (CNNs). Our approach stands out for its robustness and high accuracy.

Key Highlights:

• Novel Approach: Learn how we combine ELA with CNNs, where the output of ELA is used as input for the CNN, enhancing the detection of image forgeries.

• Innovation and Results: Discover the innovative aspects of our method and see how it achieved an impressive accuracy of 93.3% in detecting various types of image forgeries. This accuracy can be fine-tuned to produce even better results.

• Deployment: Used Streamlit to host our model and create a user interface.

Dataset Used:

We used the CASIA 2.0 dataset available on Kaggle because of its reliability, robustness, and versatility in terms of the types of image forgeries present. This is a well-balanced dataset containing over 12,000+ images.

• https://www.kaggle.com/datasets/divg07/casia-20-image-tampering-detection-dataset

Algorithms Implemented:

1. Error Level Analysis (ELA):

   • ELA reveals differences in compression levels within an image, which can detect edited or manipulated regions.
   • Benefit: Helps detect subtle differences in tampered images that might not be visible to the naked eye.

2. Convolutional Neural Networks (CNNs):

   • CNNs automatically learn visual patterns like textures and discrepancies that might indicate image forgeries.
   • Benefit: Excels at image classification tasks by learning complex features from the input data.

High-level Solution Architecture:

1. Input Image: The input image is processed using Error Level Analysis (ELA) to highlight compression inconsistencies. This gives the CNN model a better idea of which parts of the image are prone to be forged.
2. Convolutional Neural Network (CNN): The ELA output is fed into the CNN, where multiple convolutional layers extract features, and pooling layers reduce dimensions. Dropout is applied to avoid overfitting.
3. Fully Connected Layer: The features from the CNN are passed through dense layers to interpret the image’s forged or authentic characteristics.
4. Output: The system classifies the image as either forged or authentic based on the processed features.

Link to video: https://youtu.be/0iSTFComJ5I

Note: Please be aware that due to some last-minute technical issues, a part of the video has inaudible audio. We apologize for any inconvenience this may cause and appreciate your understanding.

Next Steps for Image Forgery Detection:

1. Real-Time Forgery Detection:

   • Optimize the system for real-time detection on platforms like social media using faster inference and scalable deployment.

2. Improving Accuracy:

   • Use advanced models like ResNet or Transformers and apply transfer learning to improve detection accuracy.

3. Enhanced Forgery Detection:

   • Integrate techniques for detecting sophisticated forgeries like DeepFakes to increase model robustness.

Built With

• gpu
• kaggle
• machine-learning
• python
• streamlit
• tensorflow