Zone Similarity-Based Multimodal Spam Filter for Context-Sensitive Spam Detection

As Internet usage continues to grow, the incidence of spam transmission has surged correspondingly. Even though numerous spam detection methods have been developed, most primarily rely on unimodal data such as text, and often overlook external factors like contextual relevance and multimodal data. In this paper, we propose a zone similarity-based multimodal spam filter framework which complements current methods by incorporating both contextual awareness and multimodal feature information. To the best of our knowledge, this is the first approach to leverage zone-based context-sensitive analysis into multimodal spam filter. We introduce two novel concepts, i.e., intra-modal zone similarity and intra-message zone similarity, which evaluate similarities within a single modality and across multiple modalities. These concepts enhance the peformance of machine learning-based spam filters when handling complex and multimodal spam contents. Experimental results across various model types, including traditional and CNN-enhanced spam filters, confirm the effectiveness of the proposed approach. In particular, when being applied to a BERT-CNN model, the proposed method showed the F1-score from 0.9720 to 0.9936, and the overall accuracy by as much as 0.9953. Similar performance improvements were observed across other models, consistently enhancing both precision and recall. From these empirical results, we demonstrate that the proposed approach significantly reduces false positives and false negatives, thereby enhancing the robustness and reliability of spam detection systems in multimodal data analysis.