Research map / Vision-language and multimodal / Multimodal LLMs
Neural Gate: Mitigating Privacy Risks in LVLMs via Neuron-Level Gradient Gating
Xiangkui Cao, Jie Zhang, Meina Kan, Shiguang Shan, Xilin CHEN
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 neuron-level gradient gating (Neural Gate) to editing (ablation) | MiniGPT-4 · MLLMGuard (RtA) | RtA: 0.6147 -> 0.8440 | Improved |
| add neuron-level gradient gating (Neural Gate) to editing (ablation) | MiniGPT-4 · PrivacyPair-test (EtA) | EtA: 0.9014 -> 0.9395 | Improved |
| add gate in multi-layer editing (ablation) | MiniGPT-4 · PrivacyPair-test (EtA) | EtA: 0.7391 -> 0.9213 | Improved |
| remove gate in multi-layer editing (ablation) | MiniGPT-4 · MLLMGuard (RtA) | RtA: 0.7476 -> 0.9082 | Mixed |
| add gate in single-layer editing (ablation) | LLaVA-1.5 · PrivacyPair-test (EtA) | EtA: 0.9550 -> 0.9610 | Improved |
| add gate in multi-layer editing (ablation) | LLaVA-1.5 · MLLMGuard (RtA) | RtA: 0.7614 -> 0.8715 | Improved |
| extend editing from single to multiple layers (ablation) | LLaVA-1.5 · PrivacyPair-test (EtA) | EtA: 0.9610 -> 0.9823 | Improved |
| extend editing from single to multiple layers (ablation) | MiniGPT-4 · PrivacyPair-test (EtA) | EtA: 0.9395 -> 0.9213 | Hurt |
| apply Neural Gate instead of DINM | DINM · PrivacyPair-test (EtA) | EtA: 0.9312 -> 0.9395 (MiniGPT); 0.9402 -> 0.9610 (LLaVA) | Improved |
| apply Neural Gate instead of DINM | DINM · MLLMGuard (RtA) | RtA: 0.7522 -> 0.8440 (MiniGPT); 0.6972 -> 0.7522 (LLaVA) | Mixed |
| use Neural Gate vs keyword guardrail on MLLMGuard (ablation) | keyword guardrail · MLLMGuard (RtA) | refusal rate: 17.43% -> over 75% | Improved |
| apply Neural Gate to model | LLaVA-1.5 · POPE | accuracy: 0.8513 -> 0.8556 | No clear effect |
Closest papers by meaning
- SentinelLMs: Encrypted Input Adaptation and Fine-Tuning of Language Models for Private and Secure Inference (AAAI 2024 · similarity 0.86)
- Forget to Flourish: Leveraging Machine-Unlearning on Pretrained Language Models for Privacy Leakage (AAAI 2025 · similarity 0.86)
- Privacy Backdoors: Enhancing Membership Inference through Poisoning Pre-trained Models (NeurIPS 2024 · similarity 0.86)
- Privacy Auditing of Large Language Models (ICLR 2025 · similarity 0.85)
- Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge (NeurIPS 2026 · similarity 0.85)
- Privacy-Preserving Instructions for Aligning Large Language Models (ICML 2024 · 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.
