Disentangling Visual, Semantic, and Factual Deception

A Multi-Branch Approach to Misinformation Detection

This repository contains the implementation of the manuscript "Disentangling Visual, Semantic, and Factual Deception: A Multi-Branch Approach to Misinformation Detection" (added soon).

Project Overview

The rapid dissemination of misinformation on social media, particularly through the manipulation of image-text contexts, poses a significant challenge to digital forensics. This project presents a novel multimodal architecture for detecting such misinformation by explicitly modeling the consistency between visual content, textual claims, and external factual knowledge.

Key Features:

• Deepfake Detection Branch: Assesses visual authenticity using a fine-tuned vision transformer (Deepfake Detector).
• Text-Image Consistency Branch: Leverages CLIP to measure semantic alignment between text and images.
• Encyclopedic Branch: A novel module that grounds textual claims in Knowledge Graphs (KGs) constructed dynamically from DBpedia, avoiding the noise of open-web search.

The system outputs predictions for:

1. Factual Claim Veracity (Is the text claim factually true?)
2. Visual Authenticity (Is the image real or manipulated?)
3. Cross-Modal Mismatch (Does the image support the text?)
4. Overall Truthfulness (Is the post as a whole true?)

Figure 1: Overview of the proposed multi-branch architecture combining visual, semantic, and factual deception detection modules.

Figure 2: Conversion process of textual claims into claim graph and evidence graph using DBpedia.

Results

Performance of the Global architecture on MMFakeBench and COSMOS datasets (Table 4 from manuscript).

Train-Test Data Set	Neuron	Accuracy	F1	Precision	Recall
MMFakeBench-Val	(1) Factual Claim	0.927	0.941	0.950	0.931
	(2) Real Image	0.982	0.987	0.983	0.992
	(3) Mismatch	0.823	0.718	0.740	0.698
	(3) Overall Truth	0.829	0.659	0.659	0.659
MMFakeBench-Test	(1) Factual Claim	0.983	0.988	0.986	0.990
	(2) Real Image	0.992	0.995	0.994	0.995
	(3) Mismatch	0.822	0.734	0.693	0.780
	(4) Overall Truth	0.817	0.680	0.730	0.636
MMFakeBench	(1) Factual Claim	0.926	0.946	0.919	0.974
	(2) Image Real	0.984	0.989	0.982	0.997
	(3) Mismatch	0.788	0.668	0.643	0.694
	(4) Overall Truth	0.797	0.647	0.667	0.628
COSMOS-Test	(3) Mismatch	0.948	0.947	0.947	0.947

Project Organization

├── LICENSE            
├── Makefile           
├── README.md          
├── data
│   ├── external       
│   │      ├── MMFakeBench              <- Raw test/val data from MMFakeBench renamed in train/test resp.
│   │      └── cosmos                   <- Raw COSMOS test json
│   ├── processed      
│   │      ├── MMFakeBench_graphs              <- Data for MMFFakeBench train/test/val framework
│   │      ├── MMFakeBenchOriginal_graphs      <- Data for the original MMFFakeBench test/val split
│   │      ├── MMFakeBench_graphs_stratisfied <- Data for stratified sampling experiments
│   │      ├── cosmos_formatted                <- COSMOS data set with common labeling and formatting
│   │      └── cosmos_graphs                   <- Data for COSMOS experiments
│
├── docs               
│
├── models                        
│
├── notebooks          
│
├── app                           <- Demo app
│
├── app_dev                       <- App for figures and evaluation results
│
├── pyproject.toml     
│
├── reports            
│   └── figures        
│
├── requirements.txt   
│
├── setup.cfg          
│
├── src
│
├── tests
│
└── corpus_truth_manipulation   
    │
    ├── config.py               
    └── dataset.py              

Setup

Environment

1. Ensure you have Python ~3.12 installed.
2. Install dependencies and download necessary models (e.g., spaCy) using the provided script:

chmod +x install.sh
./install.sh

Configuration (.env)

This project uses python-dotenv to manage environment variables. Create a .env file in the project root to store sensitive configuration or overrides.

touch .env

Example .env content (wandb optional, depending on your usage):

HF_TOKEN=your_huggingface_token

# Optional: Weights & Biases API Key for experiment tracking
WANDB_API_KEY=your_wandb_key
WANDB_ENTITY=your_wandb_entity

Data Preparation

The project relies on MMFakeBench and COSMOS datasets. The following commands handle formatting and graph construction. Download the raw datasets and place them in the appropriate folders under data/external/ as indicated in the project organization.

We already provide in this repo the processed train/test split for MMFakeBench-Val split in data/processed/MMFakeBenchOriginal_graphs/. It is still needed to download the raw MMFakeBench data, in data/external, for the images.

For other splits:

COSMOS Dataset

Format raw COSMOS data and build DBpedia graphs:

poetry run python src/data/process_cosmos.py --split test
poetry run python -m src.data.build_dbpedia_database --split test --dataset-folder COSMOS

MMFakeBench Dataset

Process MMFakeBench data, build knowledge graphs, and reorganize splits:

# Build graphs for all splits
poetry run python -m src.data.build_dbpedia_database --split train --dataset-folder MMFAKEBENCH
poetry run python -m src.data.build_dbpedia_database --split val --dataset-folder MMFAKEBENCH
poetry run python -m src.data.build_dbpedia_database --split test --dataset-folder MMFAKEBENCH

# Reorganize into original splits
poetry run python src/data/reorganize_mmfakebench.py

# Stratified Sampling for experiments
poetry run python -m src.data.stratified_sampling_mmfakebench

Other

Attemps to use the XFACTA dataset are visible in the codebase, but were not fully integrated due to misconfigurations in the dataset.

Training

Models can be trained using the provided SLURM scripts (for HPC environments) or locally via poetry.

Train ablation study (Global + 14 variants). Using SLURM:

sbatch hpc/train_ablation.slurm --dataset MMFAKEBENCH_val_original --no-use-validation-set

Example: Train Global model on MMFakeBench-Val (original split)

Using Poetry (Locally):

poetry run python -m src.train_mmmd \
    --n-trials 1 \
    --device gpu \
    --split train \
    --wandblog \
    --no-use-optuna-params "" "optuna-study-name" \
    --jobid "LOCAL_JOB" \
    --dataset MMFAKEBENCH_val_original \
    --no-use-validation-set

Saved model

A trained model checkpoint for the Global architecture on MMFakeBench-Test train sub-split is available on HuggingFace. This model is automatically downloaded when using the demo application in app/.

Evaluation / Testing

To test the trained models on the MMFakeBench-Val set, use the following command:

# Test ablation study models on MMFakeBench-Val set
sbatch hpc/test_ablation.slurm --dataset MMFAKEBENCH_val_original

This will train all 14 ablation models and the Global model on the MMFakeBench-Val set.

To test the model checkpoint provided on HuggingFace on the test sub-split of MMFakeBench-Val:

poetry run python -m src.test_mmmd \
    --checkpoint "models/global_trained-MMFakeBenchTest.pt/
    --studyname "test_mmfakebench-val" \
    --ablationname "global_model" \
    --jobid "0" \
    --device gpu \
    --dataset_folder MMFAKEBENCH_val_original 

Testing one model variant on COSMOS (Example):

sbatch hpc/test_global.slurm \
    --checkpoint <path_to_model_checkpoint> \
    --device "cuda:0" \
    --studyname "test_cosmos" \
    --ablationname full \
    --dataset_folder COSMOS

Demo Application

A Streamlit-based interactive demo is available to visualize the model's predictions on custom images and text.

Usage:

poetry run streamlit run app/Home.py

Figure 3: Screenshot of the demo application interface.

References

• EGMMG pipeline: Duwal, S., Shopnil, M. N. S., Tyagi, A., & Proma, A. M. (2025). Evidence-Grounded Multimodal Misinformation Detection with Attention-Based GNNs. arXiv. https://doi.org/10.48550/arXiv.2505.18221

• DeepFakeDetector on HuggingFace

• COSMOS: Aneja, S., Bregler, C., & Nießner, M. (2021). COSMOS: Catching Out-of-Context Misinformation with Self-Supervised Learning. arXiv. https://doi.org/10.48550/ARXIV.2101.06278

• MMFakeBench: Liu, X., Li, Z., Li, P., Huang, H., Xia, S., Cui, X., Huang, L., Deng, W., & He, Z. (2024). MMFakeBench: A Mixed-Source Multimodal Misinformation Detection Benchmark for LVLMs. arXiv. https://doi.org/10.48550/arXiv.2406.08772

• Knowledge Base: DBpedia.

• OpenAI-CLIP: Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., & Sutskever, I. (2021). Learning Transferable Visual Models From Natural Language Supervision. arXiv. https://doi.org/10.48550/arXiv.2103.00020