Trademark Infringement Detection in E-Commerce With Multimodality Contrastive Learning

Trademark infringement detection in e-commerce presents unique challenges due to the subtle, often multimodal nature of brand imitation. Prior approaches have predominantly focused on textual similarity or structural metadata, overlooking the visual and semantic interplay critical to modern infringement tactics. In this paper, we propose a novel multimodal framework that integrates textual and visual signals for robust trademark infringement detection. Leveraging large language models (LLMs) and vision-language models (VLMs), we introduce targeted data augmentation techniques to simulate adversarial modifications in trademark names and logos. These augmented pairs are encoded through modality-specific encoders and aligned using cross-modal and infringement-specific contrastive losses, enabling the model to learn discriminative and generalizable embeddings. Experiments on the TMID dataset demonstrate that our framework significantly outperforms text-only, visual-only, and graph-based baselines. We further evaluate model performance across infringement types and under noisy conditions, showcasing strong resilience and interpretability. Our findings highlight the promise of multimodal learning for scalable, real-world trademark protection in dynamic digital marketplaces.