Library for detecting image copy-move attack
Reason this release was yanked:
broken
Project description
This is a python package for detecting copy-move attack on a digital image.
This project is part of our paper that has been published at Springer. More detailed theories and steps are explained there.
The project is formerly written with Python 2 for our Undergraduate Thesis, which is now left unmaintained here. The original thesis is written in Indonesian that in any case can also be downloaded from here.
Description
The implementation generally manipulates overlapping blocks, and are constructed based on two algorithms:
- Duplication detection algorithm, taken from Exposing Digital Forgeries by Detecting Duplicated Image Region (alternative link); Fast and smooth attack detection algorithm on digital image using principal component analysis, but sensitive to noise and any following manipulations that are being applied after the attack phase (in which they call it post region duplication process)
- Robust detection algorithm, taken from Robust Detection of Region-Duplication Forgery in Digital Image; Relatively slower process with rough result on the detection edge but are considered robust towards noise and post region duplication process
How do we modify them?
We know that the first algorithm use coordinate
and principal_component
features, while the second algorithm use coordinate
and seven_features
.
Knowing that, we then attempt to give a tolerance by merging all the features like so:
The attributes are saved as one object. Then a lexicographical sorting is applied to the principal component and the seven features.
The principal component will bring similar block closer, while the seven features will back up the detection for a block that can't be detected by principal component due to being applied with post region duplication process (for example being blurred).
By doing so, the new algorithm will have a tolerance regarding variety of the input image. The detection result will be relatively smooth and accurate for any type of image, with a trade-off in run time as we basically run two algorithm.
Getting Started
Assuming you already have Python 3.x on your machine:
- clone this repo
- create a virtual environment and enter into it
- run
pip3 install -r requirements.txt
Example usage
API for the detection process
from pimage import copy_move
fraud_list, ground_truth_image, result_image = copy_move.detect('assets/', 'dataset_example_blur.png', 'output/', block_size=32)
fraud_list
will be the list of(x_coordinate, y_coordinate)
and the number of the blocks. If the list is not empty, we can assume that the image is being tampered. For example:
means there are 5 possible matched/identical region with 2178 overlapping blocks on each of it((-57, -123), 2178) ((-11, 140), 2178) ((-280, 114), 2178) ((-34, -305), 2178) ((-37, 148), 2178)
ground_truth_image
contains the black and white ground truth of the detection resultresult_image
is the given image where the possible fraud region will be bordered (if any)
ground_truth_image
and result_image
will be formatted as numpy.ndarray
. It can further be processed. For example, it can be exported as image like so:
import imageio
imageio.imwrite("result_image.png", result_image)
imageio.imwrite("ground_truth_image.png", ground_truth_image)
Quick command to detect an image
To quickly run the detection command for your image, the copy_move.detect_and_export()
is also provided. The command is identical with .detect()
but it also save the result to desired output path.
from pimage import copy_move
copy_move.detect_and_export('dataset_example_blur.png', 'output', block_size=32)
Determining the block_size
The first algorithm use block size of 32
pixels so this package will use the same value by default. Increasing the size means faster run time at a reduced accuracy. Analogically, decreasing the size means longer run time with increased accuracy.
Verbose mode
When running copy_move.detect()
or copy_move.detect_and_export()
, you can pass verbose=True
to output
the status of each step. The default value will be False
and nothing will be printed.
Example of the verbose output:
Processing: dataset/multi_paste/cattle_gcs500_copy_rb5.png
Step 1 of 4: Object and variable initialization
Step 2 of 4: Computing characteristic features
100%|██████████| 609/609 [04:14<00:00, 2.39it/s]
Step 3 of 4:Pairing image blocks
100%|██████████| 241163/241163 [00:00<00:00, 816659.95it/s]
Step 4 of 4: Image reconstruction
Found pair(s) of possible fraud attack:
((-57, -123), 2178)
((-11, 140), 2178)
((-280, 114), 2178)
((-34, -305), 2178)
((-37, 148), 2178)
Computing time : 254.81 second
Sorting time : 0.89 second
Analyzing time : 0.3 second
Image creation : 1.4 second
Total time : 0:04:17 second
Example image
Original image
Forgered image
Example result after detection
Another example of the result can be seen in the output
directory.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.