Q-KAN: enhancing robustness of weather removal: preprocessing-based defense against adversarial attacks
Vladimir Franc, Sos С. Agaian · 2025
Deep learning has transformed computer vision by restoring images degraded by adverse weather conditions such as rain, and haze. However, these powerful models are susceptible to adversarial attacks - subtle, often imperceptible perturbations that can severely impair performance. Traditional defense strategies frequently require extensive retraining or impose significant computational burdens, hindering practical implementation in real-world scenarios. To overcome these limitations, this paper introduces the Kolmogorov- Arnold Quaternion Convolutional Network (KAQCN), a novel, training-free defense mechanism designed to mitigate adversarial noise without altering the target model. KAQCN utilizes the Kolmogorov–Arnold theorem, a fundamental result in mathematics, to achieve universal function approximation with a considerably reduced number of parameters. This efficiency is further enhanced by deploying spline-based learnable functions, effectively capturing complex nonlinear relationships within the image data. Furthermore, KAQCN incorporates quaternion-valued convolution, which unifies the RGB channels into a single hypercomplex entity. This method preserves vital inter-channel correlations often disrupted by adversarial distortions, improving the model's robustness. Keywords: Kolmogorov-Arnold Neural Network, Quaternion Neural Network, Computer Vision, Image Processing