Detecting and Localizing Copy-Move and Image-Splicing Forgery

View on arXiv ← Back to list

Authors: Aditya Pandey, Anshuman Mitra

Published: 2022-02-08 01:14:30+00:00

AI Summary

This paper investigates methods for detecting and localizing copy-move and image-splicing forgeries in images. It compares the performance of deep learning and image transformation techniques for forgery detection and mask prediction, offering insights into their robustness.

Abstract

In the world of fake news and deepfakes, there have been an alarmingly large number of cases of images being tampered with and published in newspapers, used in court, and posted on social media for defamation purposes. Detecting these tampered images is an important task and one we try to tackle. In this paper, we focus on the methods to detect if an image has been tampered with using both Deep Learning and Image transformation methods and comparing the performances and robustness of each method. We then attempt to identify the tampered area of the image and predict the corresponding mask. Based on the results, suggestions and approaches are provided to achieve a more robust framework to detect and identify the forgeries.

Key findings

DCT-based feature extraction with a simple neural network showed superior performance in forgery detection, particularly robust to blurring. UNet models effectively localized forgeries, with improvements observed when incorporating both tampered and authentic images during training. Blurring significantly impacted the accuracy of CNN-based forgery detection models.

Approach

The authors use Error Level Analysis (ELA) and Discrete Cosine Transform (DCT) as feature extraction methods for forgery detection, followed by classification using Support Vector Machines (SVM) and neural networks. For localization, they employ a UNet architecture, experimenting with different backends and training strategies.

Datasets

CASIA2.0 dataset with corresponding ground truth masks.

Model(s)

Support Vector Classifier (SVC), Convolutional Neural Networks (CNNs), ResNet architecture, LSTM, UNet (with Conv2D and ResNet-101 backends).

Author countries

USA

← Previous