Empirical analysis of adversarial robustness in 3D Gaussian Splatting under multi-view inconsistency attacks

3D Gaussian Splatting has emerged as a promising technique for real-time novel view synthesis, achieving rendering quality comparable to neural radiance fields while enabling significantly faster inference. Despite its growing adoption in various applications including autonomous driving, robotics, and augmented reality, the adversarial robustness of 3D Gaussian Splatting remains largely unexplored. This paper presents a comprehensive empirical analysis of 3D Gaussian Splatting robustness against multi-view inconsistency attacks, which inject imperceptible perturbations into training images to disrupt the reconstruction process. We propose an adversarial attack framework that maximizes view-dependent inconsistencies while preserving visual imperceptibility through semantic-aware perturbation constraints. Our extensive experiments on the Mip-NeRF 360 dataset reveal that 3D Gaussian Splatting exhibits strong robustness to multi-view photometric inconsistency attacks, with quality degradation limited to less than one percent across standard metrics including Peak Signal-to-Noise Ratio, Structural Similarity Index, and Learned Perceptual Image Patch Similarity. We identify three key factors contributing to this robustness: multi-view averaging during optimization, the explicit Gaussian primitive representation, and the gradient-based optimization dynamics that naturally suppress view-dependent artifacts. These findings provide important insights for deploying 3D Gaussian Splatting in security-sensitive applications and suggest directions for developing more effective adversarial attack strategies.