Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Bayesian Low-Rank Posteriors for Scalable Membership Inference
Khanh Nguyen, Antonio Cinà, Andrey Barsky, Alejandra Reinares, Fabio Roli, Dimosthenis Karatzas
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a scalable membership inference attack for large VLMs that replaces explicit reference/shadow model ensembles with a Bayesian low-rank posterior approximation.
Problem
Membership inference attacks based on reference/shadow model ensembles are computationally prohibitive for large-scale vision-language models.
Approach
Replaces explicit reference models with a Bayesian low-rank posterior that estimates target-conditional score distributions, avoiding training many shadow models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace reference-model ensemble with a SWAG posterior over LoRA adapters (single fine-tuning run) | LIBRA on DeepSeek-VL2 · DocVQA multi-sample MIA, D_real | TPR@3%FPR: 14.0% vs 5.3% strongest baseline | Improved |
| Image-only setting: LIBRA vs strongest baseline | Idefics2 · DocVQA image-only MIA, D_real | TPR@3%FPR: 33.0% vs 11.3% | Improved |
| Use K=15 posterior samples instead of K=4 (ablation) | LIBRA Qwen2-VL · DocVQA multi-sample (validation) | TPR@3%FPR: 27.18 -> 31.67 | Improved |
| At matched ensemble size K=N=4, LIBRA vs LiRA (ablation) | Qwen2-VL · DocVQA multi-sample (validation) | TPR@3%FPR / AUROC: 22.17 vs 27.18 TPR; 79.42 vs 77.09 AUROC | No clear effect |
| Use SWAG posterior instead of MC Dropout (ablation) | LIBRA · DocVQA multi-sample MIA, D_real | TPR@3%FPR: DeepSeek-VL2 6.33 -> 14.00; Qwen2-VL 42.00 -> 38.33 | Mixed |
| Reference adapter mismatch (LoRA reference vs VeRA target) (ablation) | LIBRA vs LiRA/RMIA · Cross-adapter transfer, D_real | TPR@3%FPR: LIBRA 27.33 vs LiRA 20.67 vs RMIA 18.67 | Mixed |
| Increase posterior scale factor kappa from 2 to 8 (ablation) | LIBRA Qwen2-VL · Per-sample calibration | observed FPR: 12.67% -> 9% | Improved |
| Increase inference temperature beta of target model as a defense (ablation) | reference-based MIAs · Qwen2-VL DocVQA | attack strength | Hurt |
| Model reference scores with population-based quantile regression (Q-MIA) rather than reference models | Qwen2-VL DocVQA · Per-sample calibrated MIA alpha=0.03 | TPR: 7.67 vs LIBRA 61.67 | Hurt |
| Use one SWAG-based reference model instead of N=4 trained shadow models | Qwen2-VL · Training cost | training GPU-hours: 5.66 -> 1.30 | Improved |
| Apply LIBRA to sample-level Medical VQA | Qwen2-VL · SLAKE sample-level MIA | TPR@3%FPR: 3.55 -> 5.46 | Improved |
Experimental setup
Builds on: reference-model membership inference attacks
Closest papers by meaning
- Membership Inference Attacks against Large Vision-Language Models (NeurIPS 2024 · similarity 0.91)
- Black-Box Membership Inference Attack for LVLMs via Prior Knowledge-Calibrated Memory Probing (NeurIPS 2025 · similarity 0.90)
- Membership Inference Attacks against Fine-tuned Large Language Models via Self-prompt Calibration (NeurIPS 2024 · similarity 0.89)
- Vid-SME: Membership Inference Attacks against Large Video Understanding Models (NeurIPS 2025 · similarity 0.88)
- LLM Dataset Inference: Did you train on my dataset? (NeurIPS 2024 · similarity 0.87)
- DF-MIA: A Distribution-Free Membership Inference Attack on Fine-Tuned Large Language Models (AAAI 2025 · similarity 0.87)
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.
