This project leverages deep learning techniques to detect splicing forgeries in images. By utilizing Error Level Analysis (ELA) for feature extraction and a CNN-based classifier optimized with focal loss, the model effectively identifies manipulated images. It is designed for forensic applications, ensuring reliable detection of image tampering through robust evaluation metrics.
Clone the project
git clone https://github.com/Lakshit-Gupta/Image_Forgery_Detection.gitGo to the project directory
cd Image_Forgery_DetectionThe dataset and model for this project is available in the Releases section. Click the link to download the files.
If using anaconda use the following command in base(root) terminal of anaconda to resolve all the depedencies
conda env create -f tf.yaml ```conda activate tf ```Open the main_app.py and run the followind command in the Terminal
streamlit run main_app.py ```docker pull lakshitgupta/forgery_detection_image:v15.0docker run -p 8501:8501 lakshitgupta/forgery_detection_image:v15.0Then go to:
localhost:8501docker pull lakshitgupta/forgery_detection_prediction-image:v16.0docker run --gpus all -p 8501:8501 forgery_detection_prediction-image:v16.0Then go to:
localhost:8501The dataset is split into three categories:
- Train: 4498 Authentic (Class 0), 965 Tampered (Class 1)
- Validation: 964 Authentic (Class 0), 206 Tampered (Class 1)
- Test: 965 Authentic (Class 0), 208 Tampered (Class 1)
The dataset is structured as follows:
train_val_test_split_384x384/
│── train/
│ ├── au/ # Authentic Images
│ ├── tp/ # Tampered Images
│
│── val/
│ ├── au/
│ ├── tp/
│
│── test/
│ ├── au/
│ ├── tp/
This model uses transfer learning with a VGG19 backbone, pretrained on ImageNet, to extract features from images. The final classification layers are customized for splicing forgery detection.
- Feature Extraction: Uses Error Level Analysis (ELA) to highlight manipulated regions.
- Pretrained Backbone: VGG19 (transfer learning from ImageNet).
- Fully Connected Layers: Batch normalization and dropout (0.5) for regularization.
- Optimizer: Adam for stable training.
- Loss Function: Focal Loss to handle class imbalance.
Input: (384, 384, 3)
VGG19 Backbone (Pretrained on ImageNet, Transfer Learning)
Fully Connected Layers (With BatchNorm & Dropout)
Dropout (0.5)
Output: Sigmoid Activation (Binary Classification)
- Optimizer Choices: AdamW(Recommended), RMSprop, SGD with Momentum
- Loss Functions Explored:
- Binary Cross-Entropy (BCE)
- Focal Loss (for imbalanced classes)
- Dice Loss (for better foreground-background separation)
- Callbacks: Early Stopping, ReduceLROnPlateau, Model Checkpointing
The model was evaluated on the test set, achieving the following performance:
| Metric | Score |
|---|---|
| Accuracy | 96.20% |
| Precision | 89.00% |
| Recall | 89.42% |
| F1-Score | 89.21% |
The confusion matrix helps in understanding model performance in terms of True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN):
| Actual \ Predicted | ✅0 (Authentic) | ❌1 (Tampered) |
|---|---|---|
| 0 (Authentic) | 🟢 952 | 🔴24 |
| 1 (Tampered) | 🔴 16 | 🟢192 |
- 🟢True Negative (TN): Authentic image correctly classified as authentic.
- 🔴False Positive (FP): Authentic image misclassified as tampered.
- 🔴False Negative (FN): Tampered image misclassified as authentic.
- 🟢True Positive (TP): Tampered image correctly classified as tampered.
- 🖼️ Enhance noise analysis using PRNU-based techniques.
- 🛠️Integrate attention mechanisms for improved localization of forged regions.
- 🔍Explore contrastive learning for improved feature representation.
- 📡Leverage Fourier Transform Analysis for frequency-based anomaly detection.
- 🖌️Incorporate Residual Pixel Analysis (RPA) with Error Level Analysis (ELA).
- 🧩 Refine splicing detection using edge and boundary irregularities.
- 💡Improve model robustness by analyzing lighting and shadow inconsistencies.
- [Lakshit Gupta]
- [Manan Goel]
For any questions, feel free to reach out!
- CASIA 2.0 Dataset
- Research papers on image forensics