Research map / Large language models / Language, safety and interpretability
Doxing via the Lens: Revealing Location-related Privacy Leakage on Multi-modal Large Reasoning Models
Weidi Luo, Tianyu Lu, Qiming Zhang, Xiaogeng Liu, Bin Hu, Yue Zhao, Jieyu Zhao, Song Gao, Patrick McDaniel, Zhen Xiang, Chaowei Xiao
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 MLRMs (reasoning models) vs non-expert humans for geolocation | MLRMs (excluding Claude) · DoxBench | GLARE (bits): 1309.73 -> 1418.97 (Top-1 mean) | Improved |
| allow Top-3 instead of Top-1 predictions (ablation) | MLRMs on DoxBench · DoxBench | CCPA accuracy (%): 11.61% -> 14.95% | Improved |
| add clue-based CoT prompting to MLLMs on answered cases (ablation) | GPT-4.1, GPT-4o, Claude Opus 4, Claude Sonnet 4 · DoxBench answered cases, Top-1 | GLARE (bits): +137.18 bits avg | Improved |
| add clue-based CoT prompting to MLLMs on unanswered cases (ablation) | GPT-4.1, GPT-4o, Claude Opus 4, Claude Sonnet 4 · DoxBench unanswered cases, Top-1 | GLARE (bits): +1256.89 bits avg | Improved |
| add clue-based CoT prompting, Top-3 answered cases (ablation) | Claude Opus 4 · DoxBench answered cases, Top-3 | GLARE (bits): 1595.30 -> 1359.59 | Hurt |
| use GeoMiner Detector-Analyzer clue passing vs CoT baseline, answered cases Top-1 | GPT-4o and LLaMA 4 Scout · DoxBench answered cases, Top-1 | GLARE (bits): +194.31 bits avg | Improved |
| use GeoMiner vs CoT baseline, unanswered cases Top-1 | GPT-4o and LLaMA 4 Scout · DoxBench unanswered cases, Top-1 | GLARE (bits): +612.12 bits avg | Improved |
| give OpenAI o3 internet search / tools (ablation) | OpenAI o3 · Additional 50-image Level-3 US dataset, Top-1 | CCPA accuracy (%): 16.00 -> 34.00 | Improved |
| move from privacy level 1 to level 2 and level 3 images | MLRMs · DoxBench | CCPA accuracy / GLARE: Top-1 L1->L2: -11.10% CCPA, -161.77 bits | Hurt |
| mirror-reflection images instead of ordinary images | MLRMs · DoxBench Mirror class | GLARE (bits): 677.91 bits Top-1, 921.40 bits Top-3 | Hurt |
| mask a single key clue in the image (ablation) | OpenAI o3 with tools · case study (clue masking) | location precision | Mixed |
| apply Llama Guard 4 as input guard (ablation) | location privacy leakage · DoxBench defense | detection of leakage | No clear effect |
Closest papers by meaning
- Do Vision-Language Models Respect Contextual Integrity in Location Disclosure? (ICLR 2026 · similarity 0.91)
- Disrupting Hierarchical Reasoning: Adversarial Protection for Geographic Privacy in Multimodal Reasoning Models (ICLR 2026 · similarity 0.90)
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.86)
- Beyond Memorization: Violating Privacy via Inference with Large Language Models (ICLR 2024 · similarity 0.83)
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · similarity 0.83)
- Recognition through Reasoning: Reinforcing Image Geo-localization with Large 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.
