Lightweight XAI-Enhanced CNN for Fashion-MNIST Classification on Edge Devices

This paper presents a lightweight convolutional neural network (CNN) with integrated explainability for fashion item classification on resource-constrained edge devices. While deep learning has revolutionized image classification, deploying these models on edge devices remains challenging due to computational limitations. Additionally, the black-box nature of neural networks hinders trust and adoption in critical applications. We address these challenges by developing a parameter-efficient CNN architecture (296 K parameters, 1.13 MB) that achieves 91.8% accuracy on the Fashion-MNIST dataset while incorporating Gradientweighted Class Activation Mapping (Grad-CAM) for visual explanations of model decisions. Our approach demonstrates that explainability can be integrated into lightweight models without significant computational overhead, providing transparency while maintaining performance suitable for edge deployment. Experimental results show that our model achieves a favorable balance between efficiency and accuracy, with only a 0.5% accuracy reduction compared to standard CNNs despite a $\mathbf{7 5 \%}$ reduction in parameter count. The visual explanations reveal that the model appropriately focuses on discriminative features specific to each fashion category, enhancing interpretability and trust. This research contributes to the development of more transparent and resource-efficient AI systems for practical applications where both model size and explainability are critical requirements.