Geometric Curvature of Neural Decision Boundaries

45 concepts22 papers1 proposal

Deep neural networks are often viewed as "black boxes," but their behavior is fundamentally governed by the geometry of the decision boundaries they create in high-dimensional space. These boundaries are the manifolds where a model's classification changes. Recent research suggests that the local curvature and thickness of these boundaries are not just mathematical abstractions—they are the primary determinants of adversarial vulnerability, query efficiency in attacks, and the feasibility of detection.

The Locally Flat Approximation

A recurring observation in the study of neural networks is that their decision boundaries are surprisingly flat in most directions near natural data points. Liu et al. and Rahmati et al. exploit this property to design "decision-based" attacks, which only require the model's final label.

If a boundary is locally flat, an attacker can estimate the normal vector of the boundary using a small number of queries and then take a direct linear path to find the shortest adversarial perturbation. This geometric linearization allows attacks like qFool and GeoDA to outperform gradient-based methods in black-box settings, where the internal weights of the model are unknown.

While boundaries are often flat, they are not perfectly linear. In targeted attacks—where an adversary tries to force a specific incorrect label—the adversarial regions are often narrow and highly curved. In these scenarios, simple linear search methods fail because they quickly "fall off" the boundary or get trapped in local optima.

To solve this, Reza et al. and Reza et al. propose navigating the decision boundary using semicircular trajectories. Instead of moving in a straight line, the attack explores the boundary along a 2D plane, effectively "hugging" the curve of the manifold. This curvature-aware approach allows the attack to maintain its position on the boundary while iteratively reducing the distance to the original image. Similarly, Ma et al. suggest modeling the boundary as a semi-ellipsoid, which provides a more flexible geometric proxy than a flat plane or a simple sphere.

The Curvature Fingerprint for Detection

The geometry of the boundary also provides a way to distinguish natural images from adversarial ones. Fawzi et al. identified a fundamental geometric asymmetry: natural data points tend to reside on the concave side of the decision boundary. When an adversarial attack pushes an image across the boundary, it typically lands in a region where the local curvature sign is reversed, sitting on the convex side.

This "curvature fingerprint" is highly effective for detection. He et al. showed that adversarial examples inhabit regions that are geometrically "cramped"—they are much closer to boundaries in many more directions than benign samples. Tian et al. and Korkmaz et al. leverage this by measuring sensitivity to fluctuations; because adversarial points sit near volatile, high-curvature regions, they react inconsistently to small spatial transforms or loss landscape probes compared to stable, natural images.

Boundary Thickness and Robustness

If curvature is the problem, "thickness" is the solution. Khanna et al. introduced the concept of boundary thickness as a metric for robustness. A "thin" boundary is one that tilts toward the data manifold, making it easy for a tiny perturbation to cross. By using techniques like Mixup, researchers can "thicken" the buffer zone between classes, forcing the boundary to be more distant and stable.

For a more rigorous defense, Singla et al. proposed using second-order information (the Hessian) during training. By bounding the eigenvalues of the model's curvature, they can mathematically guarantee that no adversarial point exists within a certain radius. This moves the field from empirical "cat-and-mouse" games toward provable geometric security, treating the decision boundary not as a mysterious wall, but as a shape that can be measured, smoothed, and reinforced.

Key Foundational Work

Empirical Study of the Topology and Geometry of Deep Networks (2018) — contributed 3 concept elements to this direction.
Decision Boundary Analysis of Adversarial Examples (2018) — contributed 1 concept elements to this direction.
Robustness of Classifiers to Universal Perturbations: A Geometric Perspective (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_decision_boundary_curvature_f9f083d2,
  title = {decision boundary curvature},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decision-boundary-curvature/rdr_f9f083d2d255409ea985c8d83706b222},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_decision_boundary_curvature_f9f083d2_v1,
  title = {decision boundary curvature},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decision-boundary-curvature/rdr_f9f083d2d255409ea985c8d83706b222/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake