PaperlayerView exact revision on arXiv

Source-linked AI summary

Detecting and Grounding Multi-Modal Media Manipulation

Rui Shao, Tianxing Wu, Ziwei Liu

arXiv:2304.02556v1cs.CV

TL;DR

The paper addresses the gap between single-modal binary forgery detection and multimodal manipulation analysis that also grounds altered image regions and text tokens. It introduces the DGM4 dataset and HAMMER, which combines manipulation-aware unimodal alignment with cross-modal attention; experiments report superiority over baselines and show that multimodal interaction is important, especially for text grounding and detection.

  • Problem

    DGM4 addresses the lack of methods that jointly detect multimodal forgery and ground manipulated image regions and text tokens.

  • Method

    The paper builds the DGM4 dataset and proposes HAMMER, combining manipulation-aware contrastive learning with modality-aware cross-attention and hierarchical detection and grounding heads.

  • Results

    HAMMER significantly outperforms adapted CLIP and ViLT baselines across the evaluated metrics, while multimodal interaction improves performance over single-modality ablations.

  • Takeaways & Limitations

    Cross-modal correlation provides complementary information for multimodal manipulation detection and grounding, with particular importance for text-modality tasks.

Abstract

from arXiv · show

Misinformation has become a pressing issue. Fake media, in both visual and textual forms, is widespread on the web. While various deepfake detection and text fake news detection methods have been proposed, they are only designed for single-modality forgery based on binary classification, let alone analyzing and reasoning subtle forgery traces across different modalities. In this paper, we highlight a new research problem for multi-modal fake media, namely Detecting and Grounding Multi-Modal Media Manipulation (DGM^4). DGM^4 aims to not only detect the authenticity of multi-modal media, but also ground the manipulated content (i.e., image bounding boxes and text tokens), which requires deeper reasoning of multi-modal media manipulation. To support a large-scale investigation, we construct the first DGM^4 dataset, where image-text pairs are manipulated by various approaches, with rich annotation of diverse manipulations. Moreover, we propose a novel HierArchical Multi-modal Manipulation rEasoning tRansformer (HAMMER) to fully capture the fine-grained interaction between different modalities. HAMMER performs 1) manipulation-aware contrastive learning between two uni-modal encoders as shallow manipulation reasoning, and 2) modality-aware cross-attention by multi-modal aggregator as deep manipulation reasoning. Dedicated manipulation detection and grounding heads are integrated from shallow to deep levels based on the interacted multi-modal information. Finally, we build an extensive benchmark and set up rigorous evaluation metrics for this new research problem. Comprehensive experiments demonstrate the superiority of our model; several valuable observations are also revealed to facilitate future research in multi-modal media manipulation.

1. Introduction

DGM4 introduces a multimodal forgery task that jointly detects manipulation in image-text pairs and grounds manipulated image regions and text tokens. The paper contributes a large-scale dataset and HAMMER, which reasons over cross-modal interactions from shallow alignment to deep aggregation.

  • 1. Introduction: DGM4 jointly detects forgery in image-text pairs and grounds manipulated image bounding boxes and text tokens, beyond single-modal binary classification.
  • 1. Introduction: The paper constructs the first large-scale DGM4 dataset with diverse image and text manipulations and rich annotations for detection and grounding.
  • 1. Introduction: HAMMER uses manipulation-aware contrastive learning for shallow reasoning and modality-aware cross-attention for deep multimodal reasoning.

2. Related Work

Prior forgery research largely addresses single modalities or binary multimodal classification, leaving multimodal manipulation grounding insufficiently covered.

  • 2. Related Work: Deepfake detection methods mainly classify image forgeries using spatial or frequency cues, without grounding manipulation across modalities.
  • 2. Related Work: Existing multimodal misinformation datasets commonly provide small human-generated fake-news or out-of-context pairs for binary forgery detection.

3. Multi-Modal Media Manipulation Dataset

DGM4 is a large-scale human-centric news dataset combining diverse image and text manipulations with fine-grained detection and grounding annotations. Its manipulations are often spatially or token-wise subtle, creating a challenging benchmark.

  • 3. Multi-Modal Media Manipulation Dataset: DGM4 contains diverse image and text manipulation techniques with fine-grained labels supporting both manipulation detection and grounding.
  • 3. Multi-Modal Media Manipulation Dataset: The dataset targets human-centric news filtered from VisualNews image-text pairs collected from real-world news sources.
  • 3. Multi-Modal Media Manipulation Dataset: Image manipulations include face swap and face attribute editing, while text manipulations include semantic text swap and sentiment-oriented attribute editing.
  • 3.3. Dataset Statistics: The dataset comprises 230k news samples: 77,426 pristine pairs and 152,574 manipulated pairs, including 32,693 mixed-manipulation pairs.
  • 3.3. Dataset Statistics: Most manipulated image regions and text-token sets are small, making DGM4 more challenging than existing deepfake and multimodal misinformation datasets.

4. HAMMER

HAMMER hierarchically reasons about multi-modal manipulation by combining shallow manipulation-aware alignment with deeper cross-modal aggregation, then applies detection and grounding heads at multiple levels. Its losses jointly support image-text alignment, image-region grounding, token grounding, manipulation-type classification, and binary classification.

  • Architecture: HAMMER combines image and text encoders, a multi-modal aggregator, and dedicated detection and grounding heads for hierarchical manipulation reasoning.The architecture includes binary and multi-label classifiers, a bounding-box detector, and a token detector.
  • 4.2. Deep Manipulation Reasoning: The unified objective sums contrastive alignment, image grounding, manipulation-type classification, binary classification, and text grounding losses.The model also predicts four fine-grained manipulation types in addition to real/fake labels.
  • 4.1. Shallow Manipulation Reasoning: Manipulation-aware contrastive learning emphasizes inconsistency by pushing manipulated image-text pairs apart alongside unmatched pairs.The method uses momentum encoders, projection heads, and queues of recent negative image-text embeddings.
  • 4.1. Shallow Manipulation Reasoning: Cross-attention grounds manipulated image regions by producing text-informed patch embeddings and aggregating their spatial information before bounding-box prediction.The image embedding serves as query while text embeddings provide keys and values; LPAA aggregates patch features into uagg for the bbox detector.
  • 4.2. Deep Manipulation Reasoning: Deep manipulation reasoning uses modality-aware cross-attention to produce token-level embeddings that integrate textual context with image features for token grounding.The multi-modal aggregator applies multiple cross-attention layers, while momentum modules generate soft pseudo-labels to moderate token prediction under noisy text.

5. Experiments

HAMMER significantly outperforms adapted multimodal and unimodal baselines across evaluation metrics, while ablations and visualizations examine the roles of cross-modal reasoning, losses, and grounding mechanisms.

  • 5.1. Benchmark for DGM4: HAMMER significantly outperforms adapted CLIP and ViLT baselines on all evaluation metrics for multimodal manipulation detection and grounding.The result is attributed to hierarchical manipulation reasoning that models cross-modal correlations and semantic inconsistency.
  • 5.1. Benchmark for DGM4: HAMMER also exceeds unimodal deepfake and sequence-tagging methods by a large margin on single-modal detection and grounding tasks.This indicates that multimodal training transfers effectively to image and text manipulation analysis.
  • 5.2. Experimental Analysis: Removing either modality degrades performance, especially for text, showing that cross-modal interaction supplies complementary information for detection and grounding.The complete model surpasses image-only and text-only ablations in Tables 5 and 6.
  • 5.2. Experimental Analysis: Removing manipulation-aware contrastive learning causes nearly all performance to deteriorate, establishing LMAC as indispensable to the task.The loss ablation is reported in Table 7.
  • 5.2. Experimental Analysis: LPAA outperforms [CLS]-token aggregation for manipulated bounding-box grounding under every metric.This result is shown in Figure 4.
  • 5.2. Experimental Analysis: Text manipulation detection is harder than image manipulation, with text attribute manipulation identified as the hardest case.The comparison is reported in Figure 5.
  • 5.2. Experimental Analysis: Visualizations show accurate bounding-box and manipulation-type predictions for face attribute and face-swap cases, while most TS and all TA manipulated tokens are grounded.Grad-CAM further shows attention to semantically relevant image regions for manipulated text.

6. Conclusion

The paper introduces DGM4 for detecting and grounding multimodal manipulations, constructs its first large-scale dataset, and proposes HAMMER with extensive experiments demonstrating effectiveness.

  • 6. Conclusion: The paper studies DGM4, constructs the first large-scale dataset with rich annotations, and proposes HAMMER for multimodal manipulation detection and grounding.Extensive experiments are used to demonstrate the model’s effectiveness.

A. Implementation Details.

The implementation uses transformer-based image, text, and multimodal encoders together with compact MLP heads for the detection and grounding outputs.

  • A. Implementation Details.: Experiments use 8 NVIDIA V100 GPUs with PyTorch, a ViT-B/16 image encoder, and 6-layer BERT-initialized text and multimodal transformers.The text encoder and multimodal aggregator initialize from separate six-layer portions of BERTbase.
  • A. Implementation Details.: Binary, multilabel, bounding-box, and token detectors are implemented as two-layer MLPs with output dimensions 2, 4, 4, and 2, respectively.These heads provide the task-specific prediction outputs.

B. Evaluation Metrics.

The evaluation protocol covers binary and fine-grained manipulation detection alongside image-box and text-token grounding, using task-specific metrics.

  • B. Evaluation Metrics.: DGM4 is evaluated with protocols and metrics spanning binary classification, multilabel manipulation types, image bounding boxes, and text tokens.The framework is designed to assess both manipulation detection and grounding.
  • B. Evaluation Metrics.: Binary classification uses ACC, AUC, and EER, while multilabel manipulation detection uses MAP, CF1, and OF1.These metrics cover authenticity decisions and fine-grained manipulation-type recognition.
  • B. Evaluation Metrics.: Image-box grounding uses mean IoU plus IoU50 and IoU75 accuracy thresholds.IoU50 and IoU75 count predictions as correct when overlap exceeds 0.5 or 0.75.
  • B. Evaluation Metrics.: Text-token grounding uses precision, recall, and F1 score to account for the imbalance between manipulated and original tokens.The metric choice is intended to provide fairer evaluation when manipulated tokens are much rarer.
Loading 2304.02556v1…