MiniNet : A CNN for Image Forgery detection

Abstract

The exponential growth of technology has made images and videos popular digital objects. With this increase in the number of visual imagery, crimes such as Identity theft, privacy invasion, fake news, etc. have also increased. This paper presents a simple, easy-to-train, fully Convolutional Neural network, named MiniNet to detect forged images with high accuracy. The model is evaluated on existing image forgery datasets which consists of Authentic and tampered images. Our network achieved an accuracy of more than $95%$ for the 140K Real and Fake Faces $93%$ for CASIA datasets. We also experimented on various state-of-the-art (SOTA) CNN models for Transfer learning to check their performance on the given dataset. Finally, we compare the results of these models with our proposed method and demonstrate different combinations of our approach with pretrained SOTA CNN models.

Requirements

1. Python 3.5
2. Numpy 1.14.2
3. Keras 2.1.5
4. Pandas 1.3.4
5. Tensorflow 2.6.0

Dataset

1. 140K Real and Fake Faces (RFF) dataset : https://www.kaggle.com/xhlulu/140k-real-and-fake-faces
2. CASIA Dataset : https://github.com/namtpham/casia1groundtruth

Weights

To download pretrained weights for the model Trained on Deepfake : https://drive.google.com/file/d/1Ju9yR9vwwPzuBwR_ZQPeLhjte3zJlgYN/view?usp=sharing Trained on both datasets : https://drive.google.com/file/d/1w-Z5X6-c-2_9Wdp9v0QpGTlesUATPe6S/view?usp=sharing

Authors

Shobhit Tyagi, Divakar Yadav National Institute of Technology, Hamirpur, India Email :[email protected], [email protected].

Citation

Please Cite this article : Tyagi, S., Yadav, D. MiniNet: a concise CNN for image forgery detection. Evolving Systems (2022). https://doi.org/10.1007/s12530-022-09446-0