Journal of Applied Science and Engineering

 Copyright The Author(s). This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are cited.

Download Citation:  BibTeX | http://dx.doi.org/10.6180/jase.202610_33.030  

Download PDF

The meteoric expansion of social media platforms has established multi-modal data as the dominant paradigm for global digital communication. However, this evolution has concurrently facilitated the rapid dissemination of sensitive and harmful material, including hate speech, extremist propaganda, and violent imagery, which presents a severe challenge to online security and public psychological well-being. Modern automated detection systems frequently falter because they fail to bridge the heterogeneous semantic gap between visual and linguistic modalities effectively, often relying on late-stage fusion or computationally expensive Transformer architectures that struggle with long-range dependencies. Furthermore, the conventional use of multi-layer perceptrons (MLPs) as predictors often fails to capture the irregular and complex decision boundaries inherent in high-dimensional multi-modal manifolds. To address these critical limitations, we propose MamKAN, a sophisticated multi-modal fusion framework that integrates State Space Models with spline-based neural architectures. Our methodology innovates at three distinct stages: First, we implement an implicit modal interaction strategy using textual inversion, which maps visual features into pseudo-word tokens within the text embedding space to achieve robust early-stage semantic alignment. Second, we introduce a bidirectional Mamba fusion module based on state space models; this mechanism achieves linear computational complexity while simultaneously capturing comprehensive vision-to-text and text-to-vision contextual dependencies through forward and backward scanning. Finally, we replace traditional MLPs with a Kolmogorov-Arnold Network (KAN) predictor. By employing learnable B-spline activation functions on the network edges, the KAN module adaptively fits complex non-linear distributions, significantly enhancing discriminative precision. Comprehensive experiments conducted on the HMC and HarMeme benchmark datasets demonstrate that
MamKAN consistently and significantly outperforms state-of-the-art models across three metrics.

Keywords: State space models; Kolmogorov-Arnold network; Sensitive content detection