Research map / Federated learning and privacy / Differential privacy
Rethinking Visual Privacy: A Compositional Privacy Risk Framework for Severity Assessment with VLMs
Efthymios Tsaprazlis, Tiantian Feng, Anil Ramakrishna, Sai Karimireddy, Rahul Gupta, Shrikanth Narayanan
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 |
|---|---|---|---|
| provide full CPRT taxonomy in prompt (ablation) | Gemini 3 Flash · CPRT benchmark | Spearman correlation: 0.802 -> 0.872 | Improved |
| provide full CPRT taxonomy in prompt (ablation) | Gemini 3 Flash · CPRT benchmark | level accuracy: 0.403 -> 0.703 | Improved |
| provide full CPRT taxonomy in prompt (ablation) | GPT-5.2 · CPRT benchmark | MAE: 0.225 -> 0.158 | Improved |
| provide only high-level intuition about compositional risk (ablation) | Gemini 3 Flash · CPRT benchmark | Spearman correlation: 0.802 -> 0.792 | No clear effect |
| provide taxonomy in prompt to a small model (ablation) | Qwen3-VL (8B) · CPRT benchmark | Spearman correlation: 0.383 -> 0.751 | Improved |
| provide taxonomy in prompt to a small model (ablation) | Llama 3.2-VL (11B) · CPRT benchmark | Spearman correlation: 0.339 -> 0.354 | No clear effect |
| provide taxonomy in prompt to Pixtral (ablation) | Pixtral (12B) · CPRT benchmark | Spearman correlation: 0.691 -> 0.720 | No clear effect |
| add taxonomy-guided LoRA SFT | Qwen3-VL (8B) · CPRT benchmark | MAE: 0.291 -> 0.140 | Improved |
| add taxonomy-guided LoRA SFT | Qwen3-VL (8B) · CPRT benchmark | Pearson correlation: 0.636 -> 0.799 | Improved |
| add taxonomy-guided LoRA SFT | Qwen3-VL (8B) · CPRT benchmark | intra-level pairwise accuracy: 0.649 -> 0.745 | Improved |
| compare SFT 8B with frontier model on level accuracy | Qwen3-VL (8B) + SFT · CPRT benchmark | level accuracy: 0.703 -> 0.633 | Hurt |
| evaluate intra-level discrimination | VLMs · CPRT benchmark | intra-level pairwise accuracy | Hurt |
Closest papers by meaning
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.88)
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · similarity 0.87)
- PID: Prompt-Independent Data Protection Against Latent Diffusion Models (ICML 2024 · similarity 0.84)
- T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 2025 · similarity 0.84)
- Auditing Privacy Mechanisms via Label Inference Attacks (NeurIPS 2024 · 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.
