Research map / Adversarial robustness and security / Attacks on text and graphs
MrM: Black-Box Membership Inference Attacks Against Multimodal RAG Systems
Peiru Yang, Jinhua Yin, Haoran Zheng, Xueying Bai, Huili Wang, Yufei Sun, Xintian Li, Songwei Pei, Yongfeng Huang, Tao Qi
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 MrM instead of QB-MIA | Claude-3.5 RAG · Flickr | TPR@5%FPR: 16.12% -> 74.98% | Improved |
| use MrM instead of the best SB-MIA | GPT-4o-mini RAG · Flickr | AUC: 70.04% -> 80.86% | Improved |
| use MrM instead of QB-MIA | Gemini-2 RAG · COCO | TPR@5%FPR: 9.28% -> 57.18% | Improved |
| use MrM instead of the strongest baseline | GPT-4o-mini RAG · COCO | AUC: 64.42% -> 73.51% | Improved |
| add object-aware masking via object detection instead of random masking (ablation) | MrM · Flickr and COCO, 8 RAG systems | AUC and TPR@5%FPR | Improved |
| add proxy-model counterfactual mask selection (ablation) | MrM · Flickr and COCO, 8 RAG systems | AUC and TPR@5%FPR | Improved |
| use SAM2 instead of YOLO object detector (ablation) | MrM · Flickr and COCO, 8 RAG systems | AUC and TPR@5%FPR | Improved |
| increase set size K from 1 to 10 | MrM · set-level MIA | AUC: AUC close to 1.0 when K = 10 | Improved |
| add system prompt defense forbidding membership disclosure | QB-MIA · Flickr and COCO | AUC / TPR | Hurt |
| add system prompt defense forbidding membership disclosure | SB-MIA and MrM · Flickr and COCO | AUC / TPR | No clear effect |
| transform database images (flip, grayscale, crop, blur) with augmentation-aware attack | MrM · Flickr and COCO | AUC / TPR | No clear effect |
| run MrM on small open-source VLMs (7B-11B) instead of commercial models | MrM · Flickr and COCO | AUC and TPR@5%FPR | Hurt |
Closest papers by meaning
- Membership Inference Attacks against Large Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- Black-Box Membership Inference Attack for LVLMs via Prior Knowledge-Calibrated Memory Probing (NeurIPS 2025 · similarity 0.89)
- Vid-SME: Membership Inference Attacks against Large Video Understanding Models (NeurIPS 2025 · similarity 0.88)
- Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models (CVPR 2025 · similarity 0.87)
- FusionAudit: Pathway-Conditioned Robustness Auditing for Native Multimodal Models (NeurIPS 2026 · similarity 0.86)
- MRAG-Bench: Vision-Centric Evaluation for Retrieval-Augmented Multimodal Models (ICLR 2025 · 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.
