EKT-XAI: Efficient Kolmogorov-Arnold Network Transformer for Lightweight Smishing Detection with Explainable AI

Abstract

SMS phishing (smishing) attacks represent an escalating cybersecurity threat, with traditional detection approaches suffering from limited adaptability, computational inefficiency, and lack of interpretability critical for security applications. This paper introduces EKT-XAI (Efficient Kolmogorov-Arnold Network Transformer with Explainable AI), a novel framework that integrates learnable activation functions within transformer architectures to enhance SMS phishing detection while providing comprehensive model interpretability. The approach replaces traditional fixed activation functions with Kolmogorov-Arnold Network (KAN) layers that implement adaptive B-spline basis functions, enabling task-specific nonlinear transformations learned during training. The framework incorporates four complementary explainability mechanisms: attention weight visualization, LIME feature attribution, KAN activation pattern analysis, and decision path tracing, operating simultaneously with prediction to provide real-time model interpretation without computational overhead. Comprehensive evaluation on the SMS Spam Collection Dataset (5,574 messages) and SMS Phishing Dataset (5,971 messages) demonstrates exceptional performance, achieving 99.99% and 99.89% accuracy respectively, outperforming existing state-of-the-art approaches including CNN-LSTM ensembles (99.74%), BERT-based models (99.28%), and traditional ensemble methods (99.58%). Ablation studies validate the critical contribution of KAN layers (1.33% accuracy improvement) and attention mechanisms (2.77% improvement) while maintaining computational efficiency suitable for mobile deployment. The integrated explainability framework enables security analysts to understand classification decisions, validate model reasoning, and identify potential attack vectors through interpretable visualizations. The framework's computational efficiency (256-dimensional embeddings, 4 attention heads, 3 transformer layers) enables real-time inference on mobile devices while preserving privacy through on-device processing. This work establishes the first successful integration of learnable activation functions within transformer architectures for cybersecurity applications, demonstrating that adaptive neural networks combined with built-in interpretability can significantly advance mobile security capabilities while addressing practical deployment requirements for next-generation SMS protection systems.