Shedding the Facades, Connecting the Domains: Detecting Shifting Multimodal Hate Video with Test-Time Adaptation

TL;DR

This work tackles the challenge of detecting evolving hateful content in multimodal videos when training and deployment data diverge semantically. It introduces SCANNER, a layered Test-Time Adaptation framework that first aligns target samples to invariant hateful cores via centroid-guided CAN, then refines alignment with sample-adaptive SCAN, and finally prevents semantic collapse through intra-cluster diversity regularization. The resulting objective combines entropy minimization with centroid-aligned, diverse predictions, enabling robust adaptation without source data or target labels. Empirical results across three HVD benchmarks show SCANNER consistently outperforms baselines, underscoring its practical value for real-time moderation in dynamic online environments.

Abstract

Hate Video Detection (HVD) is crucial for online ecosystems. Existing methods assume identical distributions between training (source) and inference (target) data. However, hateful content often evolves into irregular and ambiguous forms to evade censorship, resulting in substantial semantic drift and rendering previously trained models ineffective. Test-Time Adaptation (TTA) offers a solution by adapting models during inference to narrow the cross-domain gap, while conventional TTA methods target mild distribution shifts and struggle with the severe semantic drift in HVD. To tackle these challenges, we propose SCANNER, the first TTA framework tailored for HVD. Motivated by the insight that, despite the evolving nature of hateful manifestations, their underlying cores remain largely invariant (i.e., targeting is still based on characteristics like gender, race, etc), we leverage these stable cores as a bridge to connect the source and target domains. Specifically, SCANNER initially reveals the stable cores from the ambiguous layout in evolving hateful content via a principled centroid-guided alignment mechanism. To alleviate the impact of outlier-like samples that are weakly correlated with centroids during the alignment process, SCANNER enhances the prior by incorporating a sample-level adaptive centroid alignment strategy, promoting more stable adaptation. Furthermore, to mitigate semantic collapse from overly uniform outputs within clusters, SCANNER introduces an intra-cluster diversity regularization that encourages the cluster-wise semantic richness. Experiments show that SCANNER outperforms all baselines, with an average gain of 4.69% in Macro-F1 over the best.

Figures (8)

• Figure 1: Examples of evolving hateful manifestations. (a) indicates more explicit forms of hateful content, (b) and (c) represent more implicit and veiled manifestations.
• Figure 2: An overview of the SCANNER framework, illustrating its progressive evolution from the base framework (CAN) to the final architecture (SCANNER). Video samples are grouped into different clusters (with two example clusters shown; different colors denote different clusters). Solid lines indicate equal weighting for all samples; dashed lines denote similarity-based adaptive weighting, with longer lines implying larger distances from the cluster core and correspondingly smaller weights.
• Figure 3: Evolution of gradient norm on CAN and SCAN frameworks during online adaptation. Source domain is MHY dataset, target domain are HMM and MHB datasets.
• Figure 4: Comparison of ground truth and predicted hateful video ratios (hateful vs. total videos) for one cluster per modality (1:visual, 2:textual, 3:audio) in SCAN. Grid regions indicate over- or under-estimation in predicted ratios relative to ground truth.
• Figure 5: Entropy comparison between modality-specific cluster (1:visual, 2:textual, 3:audio) centroids and the average entropy of target videos within corresponding cluster across two target domains.