Nip Rumors in the Bud: Retrieval-Guided Topic-Level Adaptation for Test-Time Fake News Video Detection

TL;DR

Fake News Video Detection faces rapid topic-level distribution shifts as new events emerge. The paper introduces RADAR, a retrieval-guided test-time adaptation framework that leverages stable target-domain references to adapt unstable videos on the fly without access to source data or target labels. It combines Entropy Selection-Based Retrieval, Stable Anchor-Guided Alignment, and Target-Domain Aware Self-Training to bridge cross-domain gaps and reflect evolving event distributions. Evaluations on three real-world FNVD datasets show substantial performance gains over baselines, underscoring the practical impact of retrieval-guided, source-free TTA for dynamic multimodal misinformation detection.

Abstract

Fake News Video Detection (FNVD) is critical for social stability. Existing methods typically assume consistent news topic distribution between training and test phases, failing to detect fake news videos tied to emerging events and unseen topics. To bridge this gap, we introduce RADAR, the first framework that enables test-time adaptation to unseen news videos. RADAR pioneers a new retrieval-guided adaptation paradigm that leverages stable (source-close) videos from the target domain to guide robust adaptation of semantically related but unstable instances. Specifically, we propose an Entropy Selection-Based Retrieval mechanism that provides videos with stable (low-entropy), relevant references for adaptation. We also introduce a Stable Anchor-Guided Alignment module that explicitly aligns unstable instances' representations to the source domain via distribution-level matching with their stable references, mitigating severe domain discrepancies. Finally, our novel Target-Domain Aware Self-Training paradigm can generate informative pseudo-labels augmented by stable references, capturing varying and imbalanced category distributions in the target domain and enabling RADAR to adapt to the fast-changing label distributions. Extensive experiments demonstrate that RADAR achieves superior performance for test-time FNVD, enabling strong on-the-fly adaptation to unseen fake news video topics.

Figures (11)

• Figure 1: Prediction entropy of source model on four sampled news videos from four events (E1 --- E4) on the target dataset.
• Figure 2: Concept diagram of our RADAR: the source-close news videos in the target domain are utilized to guide the robust test-time adaptation of the uncertain video instances. Instances with the same shape denote semantically relevant.
• Figure 3: Overall framework of RADAR. (1) Entropy Selection-based Retrieval mechanism provides each target news video with semantically relevant yet stable reference instances. (2) Stable Anchor-Guided Alignment module explicitly narrows cross-domain gaps for the target news videos by matching their representations with their references. (3) Target-Domain Aware Self-Training paradigm endows news videos with reference-augmented category-aware pseudo-labels for self-training.
• Figure 4: Random vs. Event-wise batch sampling.
• Figure 5: Severity of distribution shift caused by data corruptions in prior TTA methods vs. topic-level shifts in FNVD.

Theorems & Definitions (4)

• Remark 1
• Remark 2
• Remark 3
• Remark 4