Research map / Adversarial robustness and security / Attacks on text and graphs
Harnessing Hyperbolic Geometry for Harmful Prompt Detection and Sanitization
Igor Maljkovic, Maria Rosaria Briglia, Iacopo Masi, Antonio Emanuele Cinà, Fabio Roli
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use hyperbolic HSVDD on HySAC embeddings instead of Euclidean SVDD on CLIP (ablation) | HyPE · ViSU test | F1: 0.66 -> 0.98 | Improved |
| use hyperbolic HSVDD on HySAC embeddings instead of Euclidean SVDD on CLIP (ablation) | HyPE · ViSU test | precision: 0.08 -> 0.98 | Improved |
| increase nu from 0.01 to about 0.0325 (ablation) | HSVDD · ViSU validation | F1: peak F1 around 0.0325 | Improved |
| increase nu from 0.1 toward 1 (ablation) | HSVDD · ViSU validation | benign accuracy | Hurt |
| use HyPE instead of Detoxify | detection · ViSU | F1: 0.40 -> 0.98 | Improved |
| use HyPE instead of NSFW-Classifier | detection · MMA | F1: 0.75 -> 0.95 | Improved |
| use HyPE instead of GuardT2I | detection · adv-MMA | F1: 0.19 -> 0.96 | Improved |
| use HyPE instead of NSFW-Classifier on SneakyPrompt | detection · SneakyPrompt | F1: 0.78 -> 0.78 | No clear effect |
| use Word Removal instead of Thesaurus+LLM sanitization | HyPS · ViSU | fraction reclassified benign (%): 65% -> 85% | Improved |
| use Thesaurus+LLM instead of Word Removal sanitization | HyPS · ViSU | semantic similarity: 0.82 and 0.87 | Improved |
| increase adaptive attack lambda toward 1 | HyPE · ViSU adaptive attack (1,000 prompts) | detection | Hurt |
| apply HyPE filtering against Concept Combination Attack | HySAC encoder in SD · ViSU safe prompts | unsafe image occurrences: reduced to zero | Improved |
Closest papers by meaning
- VLMGuard: Bootstrapping Malicious Prompt Detectors from Unlabeled Vision-Language Prompts in the Wild (NeurIPS 2026 · similarity 0.90)
- Hyperbolic Safety-Aware Vision-Language Models (CVPR 2025 · similarity 0.89)
- FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts (AAAI 2025 · similarity 0.86)
- Attention! Your Vision Language Model Could Be Maliciously Manipulated (NeurIPS 2025 · similarity 0.86)
- Image Hijacks: Adversarial Images can Control Generative Models at Runtime (ICML 2024 · similarity 0.86)
- GuardAlign: Robust Safety Alignment in Multimodal Large Language Models (ICLR 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
