File size: 7,796 Bytes
737d93a d4f1451 e9fb869 d4f1451 e9fb869 737d93a e9fb869 393718b e9fb869 737d93a 85066e9 e9fb869 737d93a e9fb869 85066e9 e9fb869 737d93a e9fb869 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 | ---
language:
- en
license: apache-2.0
size_categories:
- 100K<n<1M
task_categories:
- token-classification
- text-classification
- image-text-to-text
- object-detection
tags:
- multimodal
- manipulation-detection
- media-forensics
- deepfake-detection
---
# SAMM: Semantic-Aligned Multimodal Manipulation Dataset
[Paper](https://huggingface.co/papers/2509.12653) | [Code](https://github.com/shen8424/SAMM-RamDG-CAP)
## Introduction
The detection and grounding of manipulated content in multimodal data has emerged as a critical challenge in media forensics. While existing benchmarks demonstrate technical progress, they suffer from misalignment artifacts that poorly reflect real-world manipulation patterns: practical attacks typically maintain semantic consistency across modalities, whereas current datasets artificially disrupt cross-modal alignment, creating easily detectable anomalies. To bridge this gap, we pioneer the detection of semantically-coordinated manipulations where visual edits are systematically paired with semantically consistent textual descriptions. Our approach begins with constructing the first Semantic-Aligned Multimodal Manipulation (SAMM) dataset.
We present **SAMM**, a large-scale dataset for Detecting and Grounding Semantic-Coordinated Multimodal Manipulation. This is the official implementation of *SAMM* and *RamDG*. We propose a realistic research scenario: detecting and grounding semantic-coordinated multimodal manipulations, and introduce a new dataset SAMM. To address this challenge, we design the RamDG framework, proposing a novel approach for detecting fake news by leveraging external knowledge.
<div align="center">
<img src='./figures/teaser.png' width='90%'>
</div>
The framework of the proposed RamDG:
<div align="center">
<img src='https://github.com/shen8424/SAMM-RamDG-CAP/blob/main/figures/RamDG.png?raw=true' width='90%'>
</div>
## Notes β οΈ
- If you want to import the CAP data into your own dataset, please refer to [this](https://github.com/shen8424/CAP).
- If you want to run RamDG on datasets other than SAMM and use CNCL to incorporate external knowledge, please ensure to configure ```idx_cap_texts``` and ```idx_cap_images``` in the dataset jsons.
- We have upgraded the SAMM JSON files. The latest versions (SAMM with CAP or without CAP) are available on July 24, 2025. Please download the newest version.
## Dataset Statistics
<div align="center">
<img src='./figures/samm_statistics.png' width='90%'>
</div>
## Annotations
```
{
"text": "Lachrymose Terri Butler, whose letter prompted Peter Dutton to cancel Troy Newman's visa, was clearly upset.",
"fake_cls": "attribute_manipulation",
"image": "emotion_jpg/65039.jpg",
"id": 13,
"fake_image_box": [
665,
249,
999,
671
],
"cap_texts": {
"Terri Butler": "Terri Butler Gender: Female, Occupation: Politician, Birth year: 1977, Main achievement: Member of Australian Parliament.",
"Peter Dutton": "Peter Dutton Gender: Male, Occupation: Politician, Birth year: 1970, Main achievement: Australian Minister for Defence."
},
"cap_images": {
"Terri Butler": "Terri Butler",
"Peter Dutton": "Peter Dutton"
},
"idx_cap_texts": [
1,
0
],
"idx_cap_images": [
1,
0
],
"fake_text_pos": [
0,
11,
13,
14,
15
]
}
```
- `image`: The relative path to the original or manipulated image.
- `text`: The original or manipulated text caption.
- `fake_cls`: Indicates the type of manipulation (e.g., forgery, editing).
- `fake_image_box`: The bounding box coordinates of the manipulated region in the image.
- `fake_text_pos`: A list of indices specifying the positions of manipulated tokens within the `text` string.
- `cap_texts`: Textual information extracted from CAP (Contextual Auxiliary Prompt) annotations.
- `cap_images`: Relative paths to visual information from CAP annotations.
- `idx_cap_texts`: A binary array where the i-th element indicates whether the i-th celebrity in `cap_texts` is tampered (1 = tampered, 0 = not tampered).
- `idx_cap_images`: A binary array where the i-th element indicates whether the i-th celebrity in `cap_images` is tampered (1 = tampered, 0 = not tampered).
## Sample Usage (Training and Testing RamDG)
The following snippets are taken from the official GitHub repository to demonstrate how to train and test the RamDG framework using this dataset.
### Dependencies and Installation
```bash
mkdir code
cd code
git clone https://github.com/shen8424/SAMM-RamDG-CAP.git
cd SAMM-RamDG-CAP
conda create -n RamDG python=3.8
conda activate RamDG
conda install --yes -c pytorch pytorch=1.10.0 torchvision==0.11.1 cudatoolkit=11.3
pip install -r requirements.txt
conda install -c conda-forge ruamel_yaml
```
### Prepare Checkpoint
Download the pre-trained model through this link: [ALBEF_4M.pth](https://storage.googleapis.com/sfr-pcl-data-research/ALBEF/ALBEF_4M.pth) and [pytorch_model.bin](https://drive.google.com/file/d/15qfsTHPB-CkEVreOyf-056JWDAVjWK3w/view?usp=sharing)[GoogleDrive].
Then put the `ALBEF_4M.pth` and `pytorch_model.bin` into `./code/SAMM-RamDG-CAP/`.
```
./
βββ code
βββ SAMM-RamDG-CAP (this github repo)
βββ configs
β βββ...
βββ dataset
β βββ...
βββ models
β βββ...
...
βββ ALBEF_4M.pth
βββ pytorch_model.bin
```
### Prepare Data
We provide two versions: SAMM with CAP information and SAMM without CAP information. If you choose SAMM with CAP information, download `people_imgs1` and `people_imgs2`, then move the data from both folders to `./code/SAMM-RamDG-CAP/SAMM_datasets/people_imgs`.
Then place the `train.json`, `val.json`, `test.json` into `./code/SAMM-RamDG-CAP/SAMM_datasets/jsons` and place `emotion_jpg`, `orig_output`, `swap_jpg` into `./code/SAMM-RamDG-CAP/SAMM_datasets`.
```
./
βββ code
βββ SAMM-RamDG-CAP (this github repo)
βββ configs
β βββ...
βββ dataset
β βββ...
βββ models
β βββ...
...
βββ SAMM_datasets
β βββ jsons
β β βββtrain.json
β β β
β β βββtest.json
β β β
β β βββval.json
β βββ people_imgs
β β βββMessi (from people_imgs1)
β β βββTrump (from people_imgs2)
β β βββ...
β β
β βββ emotion_jpg
β β
β βββ orig_output
β β
β βββ swap_jpg
βββ models
β
βββ pytorch_model.bin
```
### Training RamDG
To train RamDG on the SAMM dataset, please modify `train.sh` and then run the following commands:
```bash
bash train.sh
```
### Testing RamDG
To test RamDG on the SAMM dataset, please modify `test.sh` and then run the following commands:
```bash
bash test.sh
```
## Citation
If you find this work useful for your research, please kindly cite our paper:
```bibtex
@inproceedings{shen2025beyond,
title={Beyond Artificial Misalignment: Detecting and Grounding Semantic-Coordinated Multimodal Manipulations},
author={Shen, Jinjie and Wang, Yaxiong and Chen, Lechao and Nan, Pu and Zhong, Zhun},
booktitle={ACM Multimedia},
year={2025}
}
``` |