Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Similarity over Factuality: Are we making progress on multimodal out-of-context misinformation detection?

Stefanos-Iordanis Papadopoulos, Christos Koutlis, Symeon Papadopoulos, Panagiotis C. Petrantonakis

TL;DR

The paper tackles multimodal out-of-context misinformation detection by proposing a simple yet effective baseline, MUSE, that relies on CLIP-based multimodal similarities between verification pairs and external evidence. When combined with the Attentive Intermediate Transformer Representations (AITR), MUSE achieves state-of-the-art-like gains on NewsCLIPpings and VERITE, highlighting that architectural complexity alone does not guarantee robust OOC detection. However, the authors show that these approaches largely leverage surface-level patterns and can fail on miscaptioned or more nuanced OOC cases, calling into question current evaluation paradigms and definitions. The work advocates broader benchmarks, annotated evaluations, and expanded external-evidence strategies to advance progress beyond pattern-matching capabilities toward genuine factual and logical consistency checks.

Abstract

Out-of-context (OOC) misinformation poses a significant challenge in multimodal fact-checking, where images are paired with texts that misrepresent their original context to support false narratives. Recent research in evidence-based OOC detection has seen a trend towards increasingly complex architectures, incorporating Transformers, foundation models, and large language models. In this study, we introduce a simple yet robust baseline, which assesses MUltimodal SimilaritiEs (MUSE), specifically the similarity between image-text pairs and external image and text evidence. Our results demonstrate that MUSE, when used with conventional classifiers like Decision Tree, Random Forest, and Multilayer Perceptron, can compete with and even surpass the state-of-the-art on the NewsCLIPpings and VERITE datasets. Furthermore, integrating MUSE in our proposed "Attentive Intermediate Transformer Representations" (AITR) significantly improved performance, by 3.3% and 7.5% on NewsCLIPpings and VERITE, respectively. Nevertheless, the success of MUSE, relying on surface-level patterns and shortcuts, without examining factuality and logical inconsistencies, raises critical questions about how we define the task, construct datasets, collect external evidence and overall, how we assess progress in the field. We release our code at: https://github.com/stevejpapad/outcontext-misinfo-progress

Similarity over Factuality: Are we making progress on multimodal out-of-context misinformation detection?

TL;DR

The paper tackles multimodal out-of-context misinformation detection by proposing a simple yet effective baseline, MUSE, that relies on CLIP-based multimodal similarities between verification pairs and external evidence. When combined with the Attentive Intermediate Transformer Representations (AITR), MUSE achieves state-of-the-art-like gains on NewsCLIPpings and VERITE, highlighting that architectural complexity alone does not guarantee robust OOC detection. However, the authors show that these approaches largely leverage surface-level patterns and can fail on miscaptioned or more nuanced OOC cases, calling into question current evaluation paradigms and definitions. The work advocates broader benchmarks, annotated evaluations, and expanded external-evidence strategies to advance progress beyond pattern-matching capabilities toward genuine factual and logical consistency checks.

Abstract

Out-of-context (OOC) misinformation poses a significant challenge in multimodal fact-checking, where images are paired with texts that misrepresent their original context to support false narratives. Recent research in evidence-based OOC detection has seen a trend towards increasingly complex architectures, incorporating Transformers, foundation models, and large language models. In this study, we introduce a simple yet robust baseline, which assesses MUltimodal SimilaritiEs (MUSE), specifically the similarity between image-text pairs and external image and text evidence. Our results demonstrate that MUSE, when used with conventional classifiers like Decision Tree, Random Forest, and Multilayer Perceptron, can compete with and even surpass the state-of-the-art on the NewsCLIPpings and VERITE datasets. Furthermore, integrating MUSE in our proposed "Attentive Intermediate Transformer Representations" (AITR) significantly improved performance, by 3.3% and 7.5% on NewsCLIPpings and VERITE, respectively. Nevertheless, the success of MUSE, relying on surface-level patterns and shortcuts, without examining factuality and logical inconsistencies, raises critical questions about how we define the task, construct datasets, collect external evidence and overall, how we assess progress in the field. We release our code at: https://github.com/stevejpapad/outcontext-misinfo-progress
Paper Structure (19 sections, 3 equations, 5 figures, 5 tables)

This paper contains 19 sections, 3 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Problem Formulation
  5. Multimodal Similarities
  6. Attentive Intermediate Transformer Representations
  7. Experimental Setup
  8. Datasets
  9. External Evidence
  10. Backbone Encoder
  11. Evaluation Protocol
  12. Implementation Details
  13. Experimental Results
  14. Ablation and Comparative Studies
  15. Similarity Importance
  16. ...and 4 more sections

Figures (5)

  • Figure 1: Samples from the NewsCLIPpings dataset, along with their retrieved and re-ranked evidence and their multimodal similarities.
  • Figure 2: Overview of the (a) MUSE and (b) AITR architectures.
  • Figure 3: Performance of MUSE-MLP and MUSE-RF with limited training data.
  • Figure 4: Distributions of similarity measures on NewsCLIPpings True and OOC classes.
  • Figure 5: Distributions of similarity measures on VERITE, True vs OOC and True vs Miscaptioned classes.