Angular Gradient Sign Method: Uncovering Vulnerabilities in Hyperbolic Networks

Minsoo Jo, Dongyoon Yang, Taesup Kim · 2025

Read Angular Gradient Sign Method: Uncovering Vulnerabilities in Hyperbolic Networks by Minsoo Jo, Dongyoon Yang, Taesup Kim (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-12985.txt.

Description

Minsoo Jo, Dongyoon Yang, and Taesup Kim attack neural networks that live in curved space. Hyperbolic networks embed hierarchical data, like trees and taxonomies, into hyperbolic geometry, where they represent structure better than flat Euclidean embeddings. Standard adversarial attacks such as FGSM and PGD ignore that geometry, perturbing inputs as if the space were flat, which wastes effort and can produce inconsistent examples. The authors compute the loss gradient in the tangent space, split it into a radial part, encoding depth in the hierarchy. An angular part, encoding meaning, and perturb only along the angular direction.

The angular attack fooled models more often than conventional attacks across image classification and cross-modal retrieval tasks and several architectures. Beyond breaking models, the method reads as a diagnostic: meaning, not depth, is where hyperbolic embeddings are fragile.

Catalog record

CREATORMinsoo Jo, Dongyoon Yang, Taesup Kim
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-12985

Browse the full library · About the Tower of Babel