[Paper Review] Edited Media Understanding Frames: Reasoning About the Intent and Implications of Visual Misinformation

Jeff Da, Maxwell Forbes+5 more|arXiv (Cornell University)|Dec 8, 2020

Multimodal Machine Learning Applications40 references4 citations

TL;DR

This paper introduces Edited Media Understanding (EMU), a new vision-and-language task that requires models to reason about the intent and implications of image edits through open-ended, natural language questions. The authors present a large-scale dataset (48k QA pairs) and propose PELICAN, a multimodal model that prioritizes key image regions and uses structured reasoning, achieving human-annotated answer accuracy of 40.35%—a promising start, though significant performance gaps remain.

ABSTRACT

Multimodal disinformation, from 'deepfakes' to simple edits that deceive, is an important societal problem. Yet at the same time, the vast majority of media edits are harmless -- such as a filtered vacation photo. The difference between this example, and harmful edits that spread disinformation, is one of intent. Recognizing and describing this intent is a major challenge for today's AI systems. We present the task of Edited Media Understanding, requiring models to answer open-ended questions that capture the intent and implications of an image edit. We introduce a dataset for our task, EMU, with 48k question-answer pairs written in rich natural language. We evaluate a wide variety of vision-and-language models for our task, and introduce a new model PELICAN, which builds upon recent progress in pretrained multimodal representations. Our model obtains promising results on our dataset, with humans rating its answers as accurate 40.35% of the time. At the same time, there is still much work to be done -- humans prefer human-annotated captions 93.56% of the time -- and we provide analysis that highlights areas for further progress.

Motivation & Objective

• To address the critical gap in AI systems' ability to detect harmful intent behind image manipulations, beyond simple forgery detection.
• To develop a task that captures the nuanced intent and societal implications of image edits, moving beyond pixel-level analysis.
• To create a large-scale, natural language-based dataset of edited image pairs with rich, open-ended question-answer annotations.
• To design a multimodal model that reasons about co-reference, visual grounding, and semantic implications of edits.
• To benchmark progress on understanding visual misinformation through human evaluation and ablation studies.

Proposed method

• The task frames image editing understanding as open-ended question answering over paired source and edited images, requiring both classification and rationale generation.
• A new dataset, EMU, is constructed using Photoshop battle images to ensure diverse, semantically meaningful edits on shared source images.
• PELICAN, a vision-language model, is introduced with a directed graph of image regions to model relationships between subjects and objects.
• The model uses topological sorting of image regions and prioritizes features from annotated subjects to improve reasoning about co-reference and intent.
• It is pre-trained on Conceptual Captions and fine-tuned on EMU, with ablation studies isolating contributions of pretraining, annotated features, and graph structure.
• The model is evaluated via automatic metrics (BLEU, perplexity) and human evaluation, comparing to strong baselines including standard VLP models.

Experimental results

Research questions

• RQ1 Can vision-and-language models accurately infer the intent behind image edits, such as whether an edit is misleading or deceptive?
• RQ2 How well can models reason about the implications of edits, such as changes in social relationships or emotional tone between subjects?
• RQ3 To what extent does grounding model outputs to specific image regions (e.g., subject1, subject2) improve reasoning and reduce ambiguity?
• RQ4 How do architectural choices—like using a directed graph of image regions or including source image features—affect performance on edited image understanding?
• RQ5 What is the performance gap between state-of-the-art models and human-level understanding in interpreting the intent of image manipulations?

Key findings

• PELICAN achieves a human-rated answer accuracy of 40.35% on the EMU dataset, significantly outperforming baseline models.
• Human annotators prefer human-generated answers over model outputs 93.56% of the time, highlighting a large remaining gap in model performance.
• Ablation studies show that pretraining and annotated image features contribute meaningfully to performance, with pretraining improving accuracy by 1.93%.
• Excluding the source image has minimal impact (0.05% accuracy drop), indicating that PELICAN can function effectively with only the edited image in real-world settings.
• The model shows strong performance on surface-level understanding (e.g., detecting a gun in an image), but struggles with nuanced social and contextual interpretations.
• Perplexity is reduced by 0.86 and BLEU scores improved over baselines, confirming PELICAN’s generative quality and reasoning capability.