Online Misinformation Detection in Live Streaming Videos

Rui CaoarXiv 2025 · DOI · arXiv

Introduction

As live streaming platforms like TikTok and YouTube have grown into dominant sources of information, they have also become significant vectors for the rapid spread of misinformation. Traditional research has largely focused on "offline" detection—analyzing static text, images, or pre-recorded videos after they have already been published. However, in the context of live broadcasts, post-hoc analysis is often too late; the deceptive content has already reached a massive audience in real-time.

This paper introduces and formalizes a new research task: Online Misinformation Detection in Live Streaming Videos (MDLS). Unlike traditional methods, MDLS requires systems to identify deceptive content as a stream unfolds, emphasizing the dual requirements of accuracy and early intervention.

The MDLS Task Formulation

The authors define MDLS as a frame-level binary classification task. At any given time step $t$, the system receives a frame $F_t$, which includes both visual information ($I_t$) and audio data ($A_t$). The system must predict whether misinformation is present based solely on the current and preceding frames ($F_1, \dots, F_t$).

This formulation introduces two critical constraints that distinguish it from offline detection:

  1. Causality: The system cannot "look ahead" to future frames to understand the context of a current segment. It must make decisions based on the information available up to the present moment.
  2. Timeliness: To effectively mitigate harm, detection must occur as soon as the misinformation appears. A system that identifies a lie ten minutes after it was told is significantly less useful than one that flags it within seconds.

Proposed Methodology and Baselines

To establish a foundation for this new task, the paper proposes a framework for building MDLS baselines using existing computer vision and natural language processing tools.

Feature Extraction and Pre-processing

The framework utilizes off-the-shelf tools to convert raw video and audio into usable features. This includes generating image captions for visual frames, using pre-trained visual encoders (such as ResNet), and employing automatic speech recognition (ASR) to convert audio into text.

Multi-modal Fusion

A key insight of the paper is that misinformation often resides in the interaction between different modalities. For instance, a streamer might show a video of a standard medical procedure while providing false medical advice via audio. The authors suggest comparing uni-modal models (video-only or audio-only) against multi-modal fusion models that can detect contradictions between what is seen and what is heard.

Temporal Classification

Drawing inspiration from "Online Action Detection" in computer vision, the authors suggest using temporal classifiers like Structured Output SVM (SOSVM) or Multi-Layer Perceptrons (MLPs). These models assign scores to frames in real-time, triggering a "misinformation event" alert whenever the score exceeds a predefined threshold.

Dataset Construction and Annotation

Since MDLS is a new task, the paper outlines a methodology for constructing specialized datasets. The authors propose three primary avenues:

  • Data Adaptation: Segmenting existing misinformation video datasets into training and testing instances suitable for online evaluation.
  • Targeted Collection: Scraping trending and high-risk topics from live streaming platforms to capture real-world deceptive tactics.
  • Two-Tier Annotation: The authors advocate for a rigorous annotation process involving video-level labels (determining if the overall content is neutral or misinformation) and frame-level labels (identifying the exact start and end timestamps of deceptive segments).

Evaluation: The Importance of Timeliness

The paper argues that standard accuracy metrics are insufficient for MDLS. If a system detects misinformation only after the segment has ended, it has failed the requirement of real-time intervention. To address this, the authors propose a scoring formula that rewards early detection. This formula incorporates a discounting factor ($\gamma$), which reduces the system's score as the delay between the onset of misinformation and its detection increases.

Furthermore, the task adopts a "Neutral" baseline approach. In this setting, a video is considered neutral by default unless a specific segment of misinformation is identified, mirroring the operational reality of content moderation platforms.

Limitations and Future Directions

As a foundational proposal, the paper identifies several challenges for future research:

  • Temporal Reasoning: Current baselines often treat frames with limited context. Future models must better understand long-term sequences, such as a streamer establishing a premise in one segment and subverting it with false information later.
  • External Knowledge Integration: Misinformation detection often requires fact-checking against external databases. Future MDLS systems should aim to integrate real-time knowledge retrieval to verify claims as they are made.
  • Computational Efficiency: Performing multi-modal analysis on high-resolution streams in real-time requires significant optimization to maintain the low latency necessary for live intervention.

By formalizing the MDLS task and providing a roadmap for baseline development, this work sets the stage for more proactive and effective moderation of live digital content.

Summary updated

BibTeX
@misc{cao2025online,
  author = {Rui Cao},
  title = {{Online Misinformation Detection in Live Streaming Videos}},
  year = {2025},
  eprint = {2503.12627},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.12627},
}
Figure