Open-Domain, Content-based, Multi-modal Fact-checking of Out-of-Context Images via Online Resources

Sahar Abdelnabi; Rakibul Hasan; Mario Fritz

In brief

Out-of-context images—where a real image is paired with a false caption to spread misinformation—are a major threat. This paper proposes a multi-modal fact-checking framework that gathers visual and textual evidence from the web, then uses a Consistency-Checking Network (CCN) to verify whether image-caption pairs are consistent, achieving 84.5% accuracy versus 68.1% for baselines and rivaling human performance.

Contributions

Formalize the task of multi-modal cycle-consistency checking for out-of-context image detection, distinguishing it from prior work that checks claims against curated sources only.
Propose an automated open-domain approach that gathers both visual evidence (from web image search) and textual evidence (from web text search) without requiring curated databases.
Design the Consistency-Checking Network (CCN), a novel neural architecture that reasons over visual, textual, and multi-modal evidence using separate memory networks for visual reasoning, textual reasoning, and cross-modal consistency.
Demonstrate significant performance gains through user studies: humans achieve 65.6% accuracy without evidence, 80% with evidence, and the model achieves 84.5% on the test set.

Method

The paper tackles the problem of verifying image-caption pairs by automating the multi-modal fact-checking process. Given a query image and caption (claim), the system performs three key steps:

Evidence Gathering: The system retrieves visual evidence by searching the web for the query image using reverse image search, obtaining alternative captions and contexts. For textual evidence, it searches for text related to the caption using standard web APIs, capturing entity mentions, domains, and sources.

Consistency-Checking Network (CCN): The architecture consists of three main components:

Visual reasoning: Uses ResNet152 to embed images into feature space, then applies memory networks to reason over evidence images against the query image.
Textual reasoning: Embeds captions and text entities using sentence transformers and BERT, storing them in separate memory networks to compute consistency scores.
CLIP integration: Fine-tunes CLIP (ViT/B-32) on the task to capture image-text alignment across visual and textual evidence.

The model concatenates outputs from all components and passes them through batch normalization and a final classifier to predict pristine or falsified.

Results

On the NewsCliPpings dataset (containing 71,072 pristine and falsified out-of-context examples):

CCN achieves 84.5% accuracy, versus 68.1% for a CLIP-only baseline and 66.1% for previous approaches.
Removing visual evidence drops performance to 72.4%; removing textual evidence to 84.2%, indicating that both modalities contribute meaningfully.
User studies show humans achieve 65.6% accuracy on pristine/falsified classification without evidence; this rises to ~80% with the model's gathered evidence displayed.
The model's attention mechanisms align with human judgments about which evidence items are most relevant.

Notes

A key strength is the open-domain approach: the system does not require pre-built databases of facts or curated sources, making it scalable to new images and claims. The use of memory networks for reasoning over multiple evidence items is well-motivated and interpretable. User studies demonstrate that the retrieved evidence genuinely helps humans make better decisions, validating the approach.

Limitations include reliance on web-sourced evidence (which can itself contain misinformation), and the possibility that evidence searches may return pages with contradictory or noisy information. The paper acknowledges potential risks of automated systems in fact-checking and emphasizes the value of keeping humans in the loop.