Research map / Adversarial robustness and security / Attacks on text and graphs
TraceRouter: Robust Safety for Large Foundation Models via Path-Level Intervention
Chuancheng Shi, Shangze Li, Wenjun Lu, Wenhua Wu, Fei Shen, Cong Wang, Zifeng Cheng, Tat-Seng Chua
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 |
|---|---|---|---|
| suppress downstream causal path instead of only isolated source neurons (ablation) | TraceRouter · DMs, LLMs, MLLMs | defense success rate | Improved |
| suppress identified sensitive path rather than random non-sensitive paths (ablation) | TraceRouter · I2P nudity, SD 1.4 | defense success rate (%): 68.7 -> 99.2 | Improved |
| amplify rather than suppress sensitive paths (ablation) | TraceRouter · I2P nudity, SD 1.4 | defense success rate (%): 82.20 -> 72.20 | Hurt |
| select onset layer by first local peak instead of global maximum peak (ablation) | TraceRouter · I2P nudity, SD 1.4 | defense success rate (%): 63.69 -> 99.20 | Improved |
| select onset layer by first local peak instead of threshold (ablation) | TraceRouter · I2P nudity, SD 1.4 | defense success rate (%): 63.59 -> 99.20 | Improved |
| raise suppression factor lambda from 2 to 3 (ablation) | TraceRouter · I2P nudity / MS COCO, SD 1.4 | DSR and CLIP score: DSR 99.2 -> 99.7; CLIP 31.27 -> 29.32 | Mixed |
| raise suppression factor lambda from 0 to 2 (ablation) | TraceRouter · I2P nudity, SD 1.4 | defense success rate (%): 82.2 -> 99.2 | Improved |
| vary SAE expansion factor from 16x to 128x (ablation) | TraceRouter · I2P nudity, SD 1.4 | DSR and CLIP score: DSR 98.8 to 99.2; CLIP 31.24 to 31.28 | No clear effect |
| apply TraceRouter to LLaMA3-8B-Instruct | LLaMA3-8B-Instruct · Global-MMLU-Lite | accuracy (%): 71.30 -> 70.50 | Hurt |
| apply TraceRouter to Mistral-7B-Instruct | Mistral-7B-Instruct · Jailbreak average | defense success rate (%): 6.6 -> 98.8 | Improved |
| apply TraceRouter to FLUX.1 Dev | FLUX.1 Dev · I2P nudity | detected nudity count: 275 -> 68 | Improved |
| apply TraceRouter to LLaVA-1.5-7B vs Circuit Breakers | LLaVA-1.5-7B · MM-SafetyBench | defense success rate (%): 97.3 -> 99.6 | Improved |
Closest papers by meaning
- Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks (CVPR 2025 · similarity 0.85)
- TraceEvader: Making DeepFakes More Untraceable via Evading the Forgery Model Attribution (AAAI 2024 · similarity 0.84)
- FusionAudit: Pathway-Conditioned Robustness Auditing for Native Multimodal Models (NeurIPS 2026 · similarity 0.84)
- Diagnosing and Repairing Unsafe Channels in Vision-Language Models via Causal Discovery and Dual-Modal Safety Subspace Projection (CVPR 2026 · similarity 0.84)
- CausalDiff: Causality-Inspired Disentanglement via Diffusion Model for Adversarial Defense (NeurIPS 2024 · similarity 0.83)
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.83)
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.
