Research map / Large language models / Language, safety and interpretability
GraphShield: Graph-Theoretic Modeling of Network-Level Dynamics for Robust Jailbreak Detection
Sunghee Dong, Sungwon Yi, Kangmin Bae, Jaeyoon Kim, Seongyeop Kim
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 |
|---|---|---|---|
| concatenate features from all anchors instead of the single anchor can (ablation) | GraphShield (LLaMA-2) · JailbreakBench detection | FPR (%): 13.67 -> 7.08 | Improved |
| add anchor conditioning to the global graph (ablation) | GraphShield (LLaMA-2) · JailbreakBench detection | TPR (%): 68.25 -> 91.00 | Improved |
| add graph structure to anchor features instead of pooling (ablation) | GraphShield (LLaMA-2) · JailbreakBench detection | TPR (%): 88.82 -> 91.00 | Improved |
| add graph modeling to a cosine-only linear probe (ablation) | GraphShield (LLaMA-2) · JailbreakBench detection | TPR (%): 83.94 -> 91.00 | Improved |
| add anchor tokens to a sequence-level attention graph (ablation) | GraphShield (LLaMA-2) · JailbreakBench detection | TPR (%): 87.27 -> 91.00 | Improved |
| use default refusal anchors instead of anti-anchors (ablation) | GraphShield (LLaMA-2) · JailbreakBench detection | TPR (%): 83.70 -> 91.00 | Improved |
| concatenate all feature groups instead of individual groups (ablation) | GraphShield · JailbreakBench detection | FPR (%): LLaMA-2: 91.00% / 7.08%; Vicuna: 89.82% / 6.83% | Improved |
| augment detector with a few adaptive attack examples | GraphShield · Adaptive attacks | TPR (%): below 6% -> above 90% | Improved |
| evaluate leave-one-attack-out (unseen) instead of seen attacks (ablation) | GraphShield · JailbreakBench detection | TPR (%): JOOD TPR < 3%, Decipher about 36-37% | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GradientCuff — Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes (NeurIPS 2024) · better on 3 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.89)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.88)
- An Interpretable N-gram Perplexity Threat Model for Large Language Model Jailbreaks (ICML 2025 · similarity 0.88)
- Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape Analysis (NeurIPS 2026 · similarity 0.88)
- BlueSuffix: Reinforced Blue Teaming for Vision-Language Models Against Jailbreak Attacks (ICLR 2025 · similarity 0.88)
- MT-JailBench: A Modular Benchmark for Understanding Multi-Turn Jailbreak Attacks (NeurIPS 2026 · similarity 0.88)
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.
