Official implementation (code and trained models) of the MMM 2024 paper: "Exploring Multi-Modal Fusion for Image Manipulation Detection and Localization" (Arxiv ("submitted manuscript"): http://arxiv.org/abs/2312.01790)

Datasets

Download train datasets:

• Casiav2
• tampCOCO
• IMD2020
• FantasticReality

Download test datasets:

• Casiav1
• corel
• CocoGlide
• Columbia
• COVER
• DSO-1

The corel dataset is needed to create the Casiav1+ dataset.

The data lists for the test sets are split into manipulated and authentic images and are present in the files named:

./data/IDT-<DATASET_NAME>-manip.txt
./data/IDT-<DATASET_NAME>-auth.txt

This is intended for ease of use when evaluating for localization, where you only use manipulated images (localization F1 for authentic images is always 0!).

The data lists for training are split into train and val files for ease of use and reproducibility. We use the train-validation split proposed by Kwon et al. in CAT-Net. Train data lists are in files named:

./data/CAT-Net_splits/train/<DATASET_NAME>.txt
./data/CAT-Net_splits/val/<DATASET_NAME>.txt

For our experiments we validated on a validation split of our training datasets so the testing datasets remained completely unseen until evaluation.

Data folder structure

Then you should place the datasets in the data directory as such:

data/
├── Casiav1
│   ├── Au
│   ├── Gt
│   └── Tp
├── Casiav2
│   ├── Au
│   ├── mask
│   └── tampered
├── CocoGlide
│   ├── fake
│   ├── mask
│   └── real
├── Columbia
│   ├── 4cam_auth
│   └── 4cam_splc
├── compRAISE
│   └── <all images here>
├── corel-1k
│   ├── test_set
│   └── training_set
├── COVER
│   ├── Au
│   ├── mask
│   └── tampered
├── DSO-1
│   ├── images
│   └── masks
├── FantasticReality
│   ├── ColorFakeImages
│   ├── ColorRealImages
│   └── masks
├── IMD2020
│   ├── 1a1ogs
│   ├── 1a3oag
│       .
│       .
│       .
│   └── z41
└── tampCOCO
    └── <all images here>

Training

Preparation

Before training, you need to download the pretrained networks following the instructions here and place them in the /pretrained directory as:

pretrained/
├── segformer
├── noiseprint
└── modal_extractor

Experiment Settings

Before training, you need to create an experiment file and place it in the 'experiments' folder. This yaml file contains parameters for training. To train a model with our training settings you can use the ec_example.yaml file provided. You can however change training parameters as follows:

• You can change the learning parameters if needed or train for more epochs:

  WORKERS: 16
  ACCUMULATE_ITERS: 6
  BATCH_SIZE: 4
  WARMUP_EPOCHS: 2
  EPOCHS: 100

• You can change the optimizer or scheduler parameters:

  LEARNING_RATE: 0.005
  SGD_MOMENTUM: 0.9
  WD: 0.0005
  POLY_POWER: 0.9

• You can change the train or validation datasets:

  DATASET:
  TRAIN:
  - 'data-list-1'
  - ...
  - 'data-list-N'
  VAL:
  - 'val-data-list-1'
  - ...
  - 'val-data-list-N'

Localization Training

After that you can run an example training by:

example_train.sh

You can change the training parameters by creating a new experiment yaml file in the /experiments directory. The checkpoint is saved as:

ckpt/
└── <model_name>
    ├── best_val_loss.pth
    └── final.pth

The model_name parameter is set in the experiment yaml file as:

MODEL:
  NAME: <model_name>

Detection Training

To run detection training (phase 2) you need a localization checkpoint (produced from phase 1 training) placed in the chekpoints folder. If you want to use one of our localization checkpoints you can download them following the instructions here.

The experiment file for phase 2 training should be set for detection as:

MODEL:
  TRAIN_PHASE: 'detection'

Then you can train a model for detection and evaluate on our testing datasets as follows:

source data.sh
exp='./experiments/ec_example_phase2.yaml'
ckpt_loc='./ckpt/<path_to_localization_ckpt>'
ckpt='./ckpt/<model_name>/best_val_loss.pth'
$pint ec_train_phase2.py --ckpt $ckpt_loc --exp $exp

$pint test_detection.py --exp $exp --ckpt $ckpt --manip $columbia_manip --auth $columbia_auth
$pint test_detection.py --exp $exp --ckpt $ckpt --manip $cover_manip --auth $cover_auth
$pint test_detection.py --exp $exp --ckpt $ckpt --manip $dso1_manip --auth $dso1_auth
$pint test_detection.py --exp $exp --ckpt $ckpt --manip $cocoglide_manip --auth $cocoglide_auth
$pint test_detection.py --exp $exp --ckpt $ckpt --manip $casiav1_manip --auth $casiav1_auth

Evaluation

You can download our pretrained networks following the instructions here and place them in the /ckpt directory. Then you can evaluate a model for localization by using:

example_test.sh

and changing the relevant parameters, or more analytically:

source data.sh
exp='./experiments/ec_example.yaml'
ckpt='./ckpt/<model_name>/best_val_loss.pth'

$pint test_localization.py --exp $exp --ckpt $ckpt --manip $columbia_manip
$pint test_localization.py --exp $exp --ckpt $ckpt --manip $cover_manip
$pint test_localization.py --exp $exp --ckpt $ckpt --manip $dso1_manip
$pint test_localization.py --exp $exp --ckpt $ckpt --manip $cocoglide_manip
$pint test_localization.py --exp $exp --ckpt $ckpt --manip $casiav1_manip

In the same way you can evaluate a model for detection:

source data.sh
exp='./experiments/ec_example_phase2.yaml'
ckpt='./ckpt/<model_name>/best_val_loss.pth'

$pint test_detection.py --exp $exp --ckpt $ckpt --manip $columbia_manip --auth $columbia_auth
$pint test_detection.py --exp $exp --ckpt $ckpt --manip $cover_manip --auth $cover_auth
$pint test_detection.py --exp $exp --ckpt $ckpt --manip $dso1_manip --auth $dso1_auth
$pint test_detection.py --exp $exp --ckpt $ckpt --manip $cocoglide_manip --auth $cocoglide_auth
$pint test_detection.py --exp $exp --ckpt $ckpt --manip $casiav1_manip --auth $casiav1_auth

License

This code is provided for academic, non-commercial use only. Please also check for any restrictions applied in the code parts and datasets used here from other sources. For the materials not covered by any such restrictions, redistribution and use in source and binary forms, with or without modification, are permitted for academic non-commercial use provided that the following conditions are met:

Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation provided with the distribution.

This software is provided by the authors "as is" and any express or implied warranties, including, but not limited to, the implied warranties of merchantability and fitness for a particular purpose are disclaimed. In no event shall the authors be liable for any direct, indirect, incidental, special, exemplary, or consequential damages (including, but not limited to, procurement of substitute goods or services; loss of use, data, or profits; or business interruption) however caused and on any theory of liability, whether in contract, strict liability, or tort (including negligence or otherwise) arising in any way out of the use of this software, even if advised of the possibility of such damage.

Acknowledgements

This work was supported by the EU's Horizon 2020 research and innovation programme under grant agreement H2020-101021866 CRiTERIA.

Thanks to the public repositories:

• DELIVER
• TruFor
• CAT-Net

Citation

If you find our method useful in your work or you use some materials provided in this repo, please cite the following publication where our method and materials were presented:

@inproceedings{triaridis2024exploring,
    title={Exploring Multi-Modal Fusion for Image Manipulation Detection and Localization},
    author={Triaridis, Konstantinos and Mezaris, Vasileios},
    year={2024},
    month={Jan.-Feb.},
    booktitle={Proc. 30th Int. Conf. on MultiMedia Modeling (MMM 2024)}
}

Arxiv ("submitted manuscript") version accessible at: http://arxiv.org/abs/2312.01790