Continual Learning for Resilient Multimodal Misinformation Detection Across Sequential Crisis Events

Abstract

Misinformation evolves rapidly during crises such as pandemics, political events and natural disasters, challenging the reliability of static detection systems. Although recent multimodal deep learning approaches have integrated text and imagery to improve misinformation detection, they have assumed stationary data distributions and overlooked the sequential domain shifts that are encountered in practice. In this study, we introduce a continual multimodal misinformation benchmark that evaluates resilience across three crisis domains: COVID-19 health misinformation, socio-political misinformation, and disaster-related credibility. We evaluated fine-tuning and replay-based continual learning strategies using a CLIP-based architecture to evaluate their performance. Our results reveal severe catastrophic forgetting under naive fine-tuning, with forgetting exceeding 12\%. In contrast, experience replay nearly eliminates forgetting, improves the average task accuracy by over 12 pp, and stabilizes predictive calibration. Our findings establish continual learning as a critical component for building resilient multimodal misinformation detection systems in dynamic real-world environments.