The DGM⁴+ Dataset

Overview

DGM⁴+ is an extension of the DGM⁴ benchmark designed to support research on global scene inconsistency in multimodal misinformation.

Unlike traditional datasets that focus on local visual artifacts, DGM⁴+ emphasizes semantic plausibility between foreground, background, and textual narratives.

Figure 1: Example of global scene inconsistency in DGM⁴+.

Global Scene Inconsistency

In DGM⁴+, manipulation is introduced at the scene level.

Typical inconsistencies include:

Realistic people placed in physically impossible environments
Foreground–background mismatches
Implausible social or geopolitical contexts
Contradictory visual–textual narratives

These cases are challenging for models relying only on local cues.

Manipulation Taxonomy

Figure 2: Taxonomy of manipulations in DGM⁴+.

DGM⁴+ introduces three primary categories:

1. FG–BG (Foreground–Background Mismatch)

Image-only manipulation where the foreground is inserted into an implausible background.

2. FG–BG + Text Attribute (TA)

Global scene inconsistency combined with sentiment, stance, or affect changes in captions.

3. FG–BG + Text Swap (TS)

Global mismatch combined with irrelevant or misleading caption replacement.

These extend the original DGM⁴ local manipulations.

Dataset Composition

Here we outline the dataset build

Size

Figure 3: Dataset statistics.

Total new samples: ~5,000
Expansion over DGM⁴: +2.1%

Distribution

Category	Samples
FG–BG	2,000
FG–BG + TA	1,500
FG–BG + TS	1,500

Captions

News-style headlines
Length: 10–25 tokens
Human-readable and semantically rich

Data Generation Pipeline

The dataset is generated using a multi-stage pipeline:

Foreground extraction
Background retrieval
Controlled composition
Caption rewriting
Quality filtering
Deduplication

Quality Control

To ensure realism and diversity:

Face-count constraints (1–3 faces)
Perceptual hash deduplication
OCR-based text removal
Resolution normalization (400×256)

These steps reduce trivial detection shortcuts.

Dataset Splits

Split	Availability	Purpose
Training	Public	Model development
Validation	Public	Hyperparameter tuning
Test	Hidden	Leaderboard evaluation

License: CC BY-NC 4.0

Intended Use

DGM⁴+ is intended for:

Multimodal misinformation detection
Cross-modal reasoning
Explainable AI research
Robustness evaluation

Citation

```bibtex @article{singh2025dgm4plus, title = {DGM$^4$+: Dataset Extension for Global Scene Inconsistency}, author = {Singh, Gagandeep and Amarsinghe, Samudi and Singh, Priyanka and Li, Xue}, journal = {arXiv preprint arXiv:2509.26047}, year = {2025} }