Research map / Vision-language and multimodal / Cross-modal learning
Multi-modal Identity Extraction
Ryan Webster, Teddy Furon
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 full Wikipedia name set as distractors (S3) instead of S1 | IMIA · ViT-G-14 L2B | TPR@FPR=0.25%: 16.7 -> 34.6 | Improved |
| use full Wikipedia name set (S3) instead of S1 | IMIA · ViT-H-14 L2B | TPR@FPR=0.25%: 19.6 -> 36.4 | Improved |
| use full Wikipedia name set (S3) instead of S2 | IMIA · ViT-B-32 L2B | TPR@FPR=1%: 57.2 -> 52.8 | Hurt |
| add adaptive gap between top-2 frequencies | IESA · Convnext-xxlarge L2B | E@99.5: 0.70 -> 15.85 | Improved |
| add adaptive gap between top-2 frequencies | IESA · Evae-14 L2B | E@99.5: 0.68 -> 21.78 | Improved |
| add adaptive gap between top-2 frequencies | IESA · ViT-G-14 L2B | E@95: 52.04 -> 47.42 | Hurt |
| limit attacker to 10 images per identity instead of all (ablation) | IESA adaptive · ViT-G-14 L2B | E@99.5: 32.53 -> 18.65 | Hurt |
| limit attacker to 10 images per identity (ablation) | IESA adaptive · Evae-14 L2B | E@99.5: 21.78 -> 0.03 | Hurt |
| append 10k distractor names (IESA step) to generated names (ablation) | IEGA · VLM generations | precision | Improved |
| use GPT-4o instead of REINFORCE for name generation | IEGA · VGGFace2 4000 identities | total extraction rate: 12.68 -> 52.51 | Improved |
| use Pixtral-12B instead of REINFORCE (ViT-L-14 L2B) | IEGA · VGGFace2 4000 identities | total extraction rate: 12.68 -> 8.40 | Hurt |
| apply strict face crop to attacker images (ablation) | IESA · ViT-G-14 L2B | E@95: 44.78 -> 25.91 | Hurt |
Closest papers by meaning
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.87)
- Guard Me If You Know Me: Protecting Specific Face-Identity from Deepfakes (NeurIPS 2025 · similarity 0.86)
- Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models (CVPR 2025 · similarity 0.84)
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · similarity 0.84)
- From Weak Cues to Real Identities: Evaluating Inference-Driven De-Anonymization in LLM Agents (ICML 2026 · similarity 0.84)
- How Private is Private? A Comparative Study for Face De-Identification (NeurIPS 2026 · similarity 0.84)
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.
