TLDR: A new research paper introduces the Grounding Multimodal Misinformation (GroundMM) task, which aims to not only detect but also precisely localize misleading segments across text, speech, and video in online content. To support this, the GroundLie360 dataset is presented, featuring 2,000+ fact-checked videos with fine-grained annotations for six types of misinformation. The paper also proposes FakeMark, a VLM-based system that simultaneously classifies and grounds misinformation. Experiments reveal the complexity of GroundMM, highlighting challenges in accurately localizing various fake types, especially those involving subtle visual or cross-modal inconsistencies, and setting a new direction for explainable misinformation detection.
The digital age has brought with it an unprecedented surge in online misinformation, particularly in the form of persuasive and deceptive videos. While previous efforts have focused on simply classifying videos as ‘real’ or ‘fake,’ or localizing issues within a single modality, these approaches often fall short in providing the detailed explanations needed to truly combat sophisticated misinformation.
A new research paper, titled A New Dataset and Benchmark for Grounding Multimodal Misinformation, introduces a groundbreaking approach to this challenge. Authored by Bingjian Yang, Danni Xu, Kaipeng Niu, Wenxuan Liu, Zheng Wang, and Mohan Kankanhalli, this work defines a novel task called Grounding Multimodal Misinformation (GroundMM). GroundMM goes beyond simple detection by aiming to verify multimodal content and pinpoint the exact misleading segments across various modalities, including text, speech, and visuals.
Introducing GroundLie360: A Comprehensive Dataset
To facilitate research into GroundMM, the authors present GroundLie360, the first real-world dataset specifically designed for this task. This dataset is a significant step forward because it features a detailed taxonomy of misinformation types and offers fine-grained annotations across text, speech, and video. Crucially, all annotations are validated with evidence from Snopes, a reputable fact-checking website, and include annotator reasoning for transparency.
GroundLie360 categorizes misinformation into six distinct types:
- False Title: When the video’s headline contains demonstrably false claims.
- False Speech: When the spoken content within the video includes false claims.
- Temporal Edit: Videos altered to distort event chronologies or create deceptive narratives.
- CGI: Digitally manipulated or generated synthetic media.
- Contradictory Content: Semantic mismatches between the text and video.
- Unsupported Content: Headlines or claims that lack evidentiary support within the video content.
The dataset employs a three-level annotation scheme: first, a binary classification of veracity (Real/Fake); second, identifying the specific fake types present; and third, providing fine-grained grounding by marking the exact text span, frame indices, or bounding box where the falsehood appears. This meticulous approach ensures high data quality, with strong inter-annotator agreement scores.
FakeMark: A VLM-based Detection and Grounding System
Alongside the dataset, the paper proposes FakeMark, a Vision-Language Model (VLM)-based, question-driven baseline. FakeMark is designed as a unified framework that, given a video with its title and transcript, can simultaneously predict its veracity, identify misinformation types, and localize the fake content across modalities. It achieves this through an analysis module that performs text analysis, video temporal analysis, video spatial analysis, and cross-modal analysis.
The system then uses these analyses for binary and multilabel classification. For grounding, FakeMark includes specialized localizers: a textual localizer for false titles, false speech, contradictory, and unsupported content; and temporal edit and CGI localizers that leverage video scene segmentation and object tracking to pinpoint manipulated segments and regions within the video.
Also Read:
- Unmasking the Digital Deceivers: How Adversarial Attacks Threaten Automated Fact-Checking
- Evaluating Video Model Accuracy: Introducing MESH for Hallucination Measurement
Challenges and Future Directions
Experiments with FakeMark on the GroundLie360 dataset highlight the inherent challenges of the GroundMM task. While FakeMark offers a unified, training-free framework, its performance varies across different misinformation types. It shows better results for text-based issues like false titles but struggles with subtle visual changes in temporal edits or complex cross-modal reasoning required for contradictory content. The grounding tasks, particularly for video temporal and spatial manipulations, also present significant difficulties, indicating that direct prompting for CGI localization is still evolving.
This research marks a crucial advancement in the fight against online misinformation. By introducing the GroundMM task and the GroundLie360 dataset, the authors lay a robust foundation for developing more explainable and precise multimodal misinformation detection systems. This work is expected to benefit various applications, from automated fact-checking to general deepfake grounding and advanced cross-modal content analysis.
