Skip to content

Image forgery detection using convolutional neural networks. Group 10's final project for TU Delft's course CS4180 Deep Learning 2019.

Notifications You must be signed in to change notification settings

kPsarakis/Image-Forgery-Detection-CNN

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

🎓 Image Forgery Detection with CNNs

In this project, we used pytorch in order to implement a Convolutional Neural Network (CNN) for the purpose of extracting features in the problem of image forgery detection. This approach is inspired by the work of Y. Rao et al. A Deep Learning Approach to Detection of Splicing and Copy-Move Forgeries in Images. Following the feature fusion, proposed in the same paper, we take the extracted features and give them as input to an SVM that performs the final binary classification task. The SVM implementation was taken from scikit-learn. The datasets used in this project are the CASIA2 and the NC2016 datasets. This study was conducted as a final project of TU Delft's course CS4180 Deep Learning 2019 by Group 10.

📜 System Overview

The pipeline of the system is:

  1. Train the CNN with image patches close to the distribution of the images that the network will work on. The training patches contain both tampered and untampered regions from the corresponding images.
  2. Extract features from unseen images by breaking them into patches and applying feature fusion after the final convolutional layer of the network.
  3. Use an SVM classifier on the 400 extracted features of the previous step for the final classification.

The high-level pipeline is shown in the following image:

📐 Network Architecture

The CNN architecture of this project is shown in the image below and is influenced by the work of Y. Rao et al. The network structure is 2 convolutions, max pooling, 4 convolutions, max pooling and then 3 convolutions. In the training phase, after the final convolution, a fully connected layer with softmax is applied. In the testing phase, the 400-D output of the final convolutional layer is used in the next Feature Fusion step that creates the feature vectors.

💈 Feature Fusion

In order to create a feature representation of an image during the test phase, k patches are extracted and passed through the network. After this procedure, k 400-D feature maps are being exported. These feature maps are fused into one feature vector for each image either using max or mean fusion.

🎏 Classification with SVM

For the final part of the pipeline an SVM classifier is trained and tested using the 400-D representations from the previous step. In particular, we use stratified 10-fold cross-validation to obtain an unbiased error estimate.

📊 Results

The accuracy and cross-entropy loss per epoch during the CNN training for the two datasets is shown below:

The SVM classification accuracy on both datasets after the 10-fold cross-validation is presented in the table below:

Dataset Accuracy
CASIA2 96.82% ± 1.19%
NC2016 84.89% ± 6.06%

For more detailed information feel free to take a look at our project report.

🏢 Project Structure Codacy Badge

The structure of the project is:

  • data Here lay all the data files related to the project. The CASIA2 and NC16 folders are empty because GitHub does not allow files of such size.

    • output In this folder we have all the outputs of the pipeline.
      • accuracy CSVs containing the accuracy per epoch in all our runs.
      • features CSVs containing the final feature representations of every image after the feature fusion part. To minimize the repo size we only maintained two feature files (one per dataset) as an example.
      • loss_function CSVs containing the loss per epoch in all our runs.
      • pre_trained_cnn Pt files that contain the trained CNNs of all our runs.
  • reports Final report of the project that contains more details on the implementation.

  • src Source folder of the project. Here we give examples on how to run every part of the pipeline.

    • classification Folder containing the SVM code.
    • cnn Folder containing the CNN code.
    • feature_fusion Folder containing the code used for the feature fusion.
    • patch_extraction Folder containing the code used for the patch extraction.
    • plots Folder containing the code used for the plots that we generated.

👥 Group 10 Team Members

Achilleas Vlogiaris

Arkajit Bhattacharya

Kyriakos Psarakis

Panagiotis Soilis

Rafail Skoulos

About

Image forgery detection using convolutional neural networks. Group 10's final project for TU Delft's course CS4180 Deep Learning 2019.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages