Adversarial Sensitivity-Augmented Graph Neural Networks for Robust Inverse Drainage Identification

Abstract

This paper introduces Adversarial Sensitivity-Augmented Graph Neural Networks (ASA-GNNs), a principled framework designed to tackle the challenges of inverse drainage identification when input data is noisy or contaminated by outliers. Inverse drainage problems are, by their very nature, ill-posed — small perturbations in the input can lead to dramatically different inferred outputs. Standard graph neural networks tend to falter under such conditions, particularly when deployed on real-world hydrological data that is rarely clean or complete. Our approach directly confronts these difficulties by weaving adversarial training principles into the sensitivity-augmented message passing process.
At the heart of ASA-GNN lies the Sensitivity-Augmented Message Passing (SAMP) layer. Unlike conventional aggregation schemes, SAMP dynamically weights incoming messages according to loss sensitivity gradients and simultaneously employs distance-aware label smoothing to bring predictions for clean and perturbed inputs into alignment. We complement this with a gradient diversity regularization term that discourages inconsistent parameter updates across different adversarial variants of the same input, thereby improving generalization. A proximal optimization strategy ties these components together, keeping the training process stable even under aggressive adversarial perturbations.
Through extensive experiments, ASA-GNN consistently outperforms competing methods in both accuracy and robustness — especially in scenarios involving missing or spurious hydrological connections. Crucially, the framework uses a graph transformer backbone and adaptive adversarial attacks during training but requires no architectural changes at inference time, making it straightforward to deploy. Beyond drainage modeling, the ideas developed here carry broader implications for robust graph learning in urban planning, environmental monitoring, and related infrastructure domains.