Abstract
Adversarial vulnerability remains a significant hurdle in the deployment of reliable neural networks. While traditional defenses focus on optimization-based adversarial training, the structural reasons for model fragility are less understood. This paper introduces Neural Ricci Curvature (NRC), a geometric metric designed to identify "bottleneck" edges within a network's architecture. By treating neural networks as data-dependent graphs, the authors demonstrate that adversarial vulnerability is closely linked to the presence of edges with negative curvature—structural "choke points" that force data through narrow pathways. Our analysis shows that robust models naturally possess a more balanced internal geometry, suggesting that network science offers a powerful diagnostic tool for understanding and improving model reliability.
Introduction
Neural networks (NNs) are notoriously fragile. A tiny, imperceptible change to an input—an adversarial perturbation—can cause a state-of-the-art model to confidently misclassify an image. Current defenses, such as adversarial training, typically treat this as a minimax optimization problem: minimizing loss against a hypothetical worst-case attacker. However, this approach provides little insight into the underlying structural causes of vulnerability.
This research shifts the focus from pure optimization to network science. By asking whether the "shape" or "geometry" of data flow influences robustness, the authors introduce a method to identify structural bottlenecks. This geometric perspective allows us to move beyond observing failure to diagnosing the specific architectural features that make a model susceptible to attack.
Methodology: The Geometry of Data Flow
Graph Curvature and Bottlenecks
In differential geometry, Ricci curvature describes how shapes deform. When applied to graph theory, Ollivier-Ricci Curvature (ORC) measures the connectivity between neighboring nodes.
- Positive Curvature: Indicates a well-connected region where mass (data) can move between nodes via multiple redundant paths.
- Negative Curvature: Indicates a "bridge" or bottleneck where mass is forced through a single edge to reach its destination.
The Neural Data Graph
To apply these concepts to deep learning, the authors construct a Neural Data Graph that represents how a specific input $x$ propagates through the layers. This is not a static representation of weights but a dynamic, data-dependent graph:
- ReLU Handling: If a neuron is inactive for a specific input, its corresponding edges are removed from the graph.
- Mixed-Sign Weights Normalization: Because standard graph theory requires positive weights (representing conductance), the authors developed an algorithm to normalize neural weights, ensuring they accurately reflect the flow of information regardless of their original sign.
By calculating the NRC for every edge in this graph, the researchers can pinpoint exactly where a specific input encounters structural resistance within the model.
Experimental Setup
The study utilizes the MNIST dataset, a benchmark for studying fundamental properties of adversarial robustness. The researchers evaluated several architectures:
- Fully Connected (FC) Networks: Small [15, 20] and medium [15, 25, 20, 15] configurations.
- Convolutional Neural Networks (CNNs): A standard architecture comprising two convolutional layers followed by two FC layers.
Models were trained using three distinct strategies: standard Cross-Entropy (CE), Weight Decay (WD), and Adversarial Training (AT). This allowed for a direct comparison between standard models and those specifically engineered for robustness.
Key Insights and Results
Bottlenecks as a Proxy for Fragility
The central finding of the study is the strong correlation between negative curvature and adversarial fragility. In non-robust models, data is often forced through a small number of "bridge" edges. A minor perturbation at the input can drastically shift the signal across these bridges, leading to a cascade of errors and a final misclassification.
Adversarial Training and Geometric Smoothing
Analysis of the Cumulative Distribution Function (CDF) of NRC values revealed that models trained via Adversarial Training (AT) exhibit significantly fewer negative-curvature edges. In the [15, 20] architecture, for instance, standard training resulted in near-zero robust accuracy at $\epsilon=0.2$. In contrast, the AT model maintained 27% accuracy and showed a much higher Area Under the Curve (AUC) for its curvature distribution. This suggests that adversarial training works by structurally "widening" the paths through the network, effectively eliminating the geometric bottlenecks that attackers exploit.
Consistency Across Architectures
This trend—where more negative curvature equals less robustness—held true across various network depths. While the effect was slightly less pronounced in CNNs due to their inherently sparse connectivity, the fundamental relationship between geometric bottlenecks and vulnerability remained consistent.
Limitations and Future Work
While NRC provides a powerful diagnostic tool, it faces challenges in scalability. Calculating Ricci curvature involves solving an "Optimal Transport" problem (the Wasserstein distance) for every edge, which is computationally expensive for massive models like ResNets or Transformers. Future research will focus on developing efficient approximations to allow for the analysis of high-dimensional datasets like ImageNet.
Furthermore, the authors propose moving from diagnosis to treatment. By incorporating NRC as a regularization term during training, it may be possible to penalize negative curvature directly. This would force the network to develop a robust, "well-curved" internal geometry from the outset, potentially creating models that are inherently resistant to adversarial manipulation without the need for exhaustive adversarial training cycles.
Conclusion
By bridging the gap between differential geometry and deep learning, this work provides a new lens through which to view neural network robustness. The identification of negative curvature as a marker for vulnerability suggests that the "shape" of a network is just as important as its weights. As we move toward more reliable AI, understanding and smoothing these geometric bottlenecks will be a critical step in building models that are robust by design.
Summary updated