Detecting and Grounding Multi-Modal Media Manipulation

Rui Shao; Tianxing Wu; Ziwei Liu

In brief

Existing deepfake and text fake-news detection methods operate on single modalities, missing manipulation traces that cross image–text boundaries. This paper introduces DGM¹, a novel problem of detecting and localizing media manipulation in image-text pairs, contributes a large-scale dataset with diverse manipulations, and proposes HAMMER, a hierarchical multi-modal model that outperforms single-modality approaches by reasoning over cross-modal inconsistencies.

Contributions

Introduces DGM¹ (Detecting and Grounding Multi-Modal Media Manipulation), a novel research problem that detects both the presence and location of manipulations in multi-modal content.
Constructs the first large-scale DGM¹ dataset with 77,426 pristine and 152,574 manipulated image-text pairs, annotated with fine-grained manipulation types and ground-truth bounding boxes and text tokens.
Proposes HAMMER (HierArchical Multi-modal Manipulation rEasoning), a powerful model that aligns image and text embeddings through manipulation-aware contrastive learning and aggregates multi-modal information hierarchically for detection and grounding.
Demonstrates extensive experiments showing HAMMER's superiority over existing single-modal and multi-modal learning baselines for manipulation detection and grounding.

Method

Dataset Construction. The authors collect image-text pairs from real-world news sources (The Guardian, BBC, USA TODAY, Washington Post) and apply two categories of manipulations:

Swap-type manipulations: Face swap (identity substitution via GAN-based methods) and text swap (altering semantic content while preserving character identity).
Attribute-type manipulations: Face attribute manipulation (e.g., changing emotion, smile) and text attribute manipulation (e.g., adding sentiment bias to modify tone).

Manipulated pairs are annotated with binary labels, fine-grained manipulation type labels, and spatial grounding (bounding boxes for image regions, token-level labels for text).

HAMMER Architecture. The model consists of:

Shallow Manipulation Reasoning: Image and text encoders (ViT-B/16 and BERT-based) extract embeddings. Manipulation-aware contrastive learning aligns embeddings between original and manipulated pairs, emphasizing semantic inconsistencies caused by manipulation.
Deep Manipulation Reasoning: A Multi-Modal Aggregator $F$ learns hierarchical embeddings by cross-attention between image and text. Dedicated manipulation detection heads predict binary classification (real/fake) and fine-grained manipulation type. Token-level grounding uses cross-attention between image patches and text tokens to localize manipulated regions.

Loss Functions. The model optimizes:

Manipulation-aware contrastive loss $\mathcal{L}_{\text{MAC}}$ for semantic alignment
Binary classification loss $\mathcal{L}_{\text{BIC}}$
Fine-grained manipulation-type loss $\mathcal{L}_{\text{MLC}}$
Image bounding-box grounding loss $\mathcal{L}_{\text{IMG}}$
Text token grounding loss $\mathcal{L}_{\text{TMG}}$

Results

On the DGM¹ dataset:

HAMMER achieves 86.39% accuracy for binary classification, substantially outperforming single-stream baselines (CLIP: 76.40%, ViLT: 78.38%) and dual-stream approaches adapted from the literature.
Multi-label fine-grained manipulation-type classification: 80.37% accuracy, compared to 52.31% for CLIP and 60.09% for ViLT.
Image bounding-box grounding: 83.75% IoU (Intersection over Union), substantially exceeding single-modal baselines.
Text token grounding: 71.35% F1-score for identifying manipulated tokens.

Ablation studies confirm that manipulation-aware contrastive learning and hierarchical multi-modal aggregation are both essential; removing either substantially degrades performance across all tasks.

Notes

Strengths: The hierarchical reasoning approach elegantly captures both shallow inconsistencies (e.g., expression mismatch) and deep semantic correlations. The DGM¹ dataset's fine-grained annotations (manipulation type, localization) enable richer evaluation than binary detection. Extensive ablations rigorously validate design choices.

Limitations and Open Questions: The dataset focuses on political news; generalization to other domains (entertainment, product photos) is unexplored. Temporal manipulations (video frame sequences) and audio-visual modalities remain beyond scope. Robustness to adversarial perturbations on image-text pairs is not evaluated. The token-level text grounding is still relatively coarse (word-level, not character-level), which may miss subtle edits.

Impact: This work establishes a new research direction by identifying cross-modal manipulation as a distinct problem and providing both the dataset and baseline model for future work. The hierarchical reasoning framework generalizes to other multi-modal forgery scenarios.