Geometry-Aware Adversarial Attacks Target Contrastive Embedding Manifolds
A recent paper published on arXiv (2608.10237) presents a framework for adversarial attacks that is informed by geometry, specifically tailored for contrastive learning and Siamese embedding models. These models, essential for contemporary verification systems, rely on relational geometry within embedding space instead of fixed classification boundaries. The authors contend that current adversarial attacks focus too heavily on classification, neglecting this aspect of vulnerability. Their innovative approach redefines attacks as manifold-level relational corruption, which disrupts similarity organization by separating positive pairs and bringing negative pairs closer, ultimately distorting the pairwise similarity structure. To facilitate scalable application, the method transitions iterative online optimization into an offline adversarial geometry deformation prior. This paper is titled 'Beyond Decision Boundaries: Relational Geometry Attacks on Contrastive Embedding Manifolds.'
Key facts
- Paper ID: arXiv:2608.10237
- Title: 'Beyond Decision Boundaries: Relational Geometry Attacks on Contrastive Embedding Manifolds'
- Introduces a geometry-aware adversarial attack framework for contrastive learning and Siamese embedding models
- Targets relational geometry in embedding space rather than classification boundaries
- Method pushes positive pairs apart and pulls negative pairs closer to collapse similarity structure
- Uses offline adversarial geometry deformation prior for scalable deployment
- Published on arXiv as a preprint
- Focuses on vulnerability of verification systems based on contrastive models
Entities
Institutions
- arXiv