Research map / Adversarial robustness and security / Attacks on text and graphs
When Visual Privacy Protection Meets Multimodal Large Language Models
Xiaofei Hui, Qian Wu, Haoxuan Qu, Majid Mirmehdi, Hossein Rahmani, Jun Liu
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 augmented Tchebycheff objective instead of weighted sum (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | Action Top-1: 32.9 -> 48.4 | Improved |
| use augmented Tchebycheff objective instead of weighted sum (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | Privacy cMAP (lower better): 58.4 -> 54.6 | Improved |
| add augmentation term to the Tchebycheff norm (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | Action Top-1: 42.6 -> 48.4 | Improved |
| add augmentation term to the Tchebycheff norm (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | Privacy cMAP (lower better): 57.3 -> 54.6 | Improved |
| add history gradients to SPSA estimation (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | Action Top-1: 35.7 -> 48.4 | Improved |
| add history gradients to SPSA estimation (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | Privacy cMAP (lower better): 57.5 -> 54.6 | Improved |
| add critical gradient collection by L2 norm (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | Action Top-1: 44.2 -> 48.4 | Improved |
| add gradient direction improvement with history (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | Action Top-1: 43.4 -> 48.4 | Improved |
| add adaptive step size adjustment (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | Action Top-1: 43.9 -> 48.4 | Improved |
| use critical-history optimization instead of SPSA (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | MLLM calls per sample: 136 -> 24 | Improved |
| use critical-history optimization instead of SPSA-GC (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | Action Top-1: 37.9 -> 48.4 | Improved |
| use critical-history optimization instead of evolution strategies (ablation) | Video-LLaVA anonymizer · UCF101-VISPR | Action Top-1: 30.2 -> 48.4 | Improved |
Closest papers by meaning
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.88)
- MultiTrust: A Comprehensive Benchmark Towards Trustworthy Multimodal Large Language Models (NeurIPS 2024 · similarity 0.87)
- MLLMGuard: A Multi-dimensional Safety Evaluation Suite for Multimodal Large Language Models (NeurIPS 2024 · similarity 0.86)
- PID: Prompt-Independent Data Protection Against Latent Diffusion Models (ICML 2024 · similarity 0.86)
- Gradient Inversion of Multimodal Models (ICML 2025 · similarity 0.86)
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · similarity 0.86)
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.
