Research map / Adversarial robustness and security / Attacks on text and graphs
EigenShield: Inference-Time, Model-Agnostic Jailbreaking Defense via Causal Subspace Filtering
Nastaran Darabi, Devashri Naik, Sina Tayebati, Dinithi Jayasuriya, Ranganath Krishnan, Amit Ranjan Trivedi
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 |
|---|---|---|---|
| add EigenShield causal-subspace projection | GPT-2 · Jailbreak prompts | attack success rate: 82.34 -> 5.81 | Improved |
| add EigenShield causal-subspace projection | LLaVA-v1.5-7B · Jailbreak prompts | attack success rate: 81.61 -> 19.20 | Improved |
| add EigenShield under adaptive AutoBreach | Llama-3.1 · AutoBreach | attack success rate: 89.53 -> 21.06 | Improved |
| add EigenShield on OOD harmful inputs | MiniGPT-4 · OOD-style harmful inputs | attack success rate: 60.31 -> 11.88 | Improved |
| increase overlap between causal and adversarial subspaces (pi/2 to pi/6) (ablation) | EigenShield · synthetic spiked model | post-filter ASR: 5% -> 45% | Hurt |
| use EigenShield vs UNIGUARD on LLaVA | EigenShield · Jailbreak prompts | attack success rate: 25.17 -> 19.20 | Improved |
| add EigenShield to Llama-3 on latency | Llama-3 · latency | latency (ms): 403.8 -> 480.1 | Hurt |
| apply EigenShield to clean data | LLMs and VLMs · clean RTP | toxicity reduction: 0.0-2.0 points | No clear effect |
| raise coverage gamma (ablation) | EigenShield · LLaVA-v1.5, MiniGPT-4 | attack success rate | Improved |
Closest papers by meaning
- DataFreeShield: Defending Adversarial Attacks without Training Data (ICML 2024 · similarity 0.86)
- Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks (CVPR 2025 · similarity 0.86)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.86)
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.86)
- VERA-V: Variational Inference Framework for Jailbreaking Vision-Language Models (ICML 2026 · similarity 0.86)
- AutoAdvExBench: Benchmarking Autonomous Exploitation of Adversarial Example Defenses (ICML 2025 · 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.
