SoK: Data Reconstruction Attacks Against Machine Learning Models...
summary
Provides a formal definition and taxonomy for vision-domain data reconstruction, plus a benchmark that evaluates 10 attacks on models trained with 100–20,000 samples. Proposes dataset-level FID and sample-level similarity plus coverage alpha to capture quantifiability, consistency, precision and diversity. Empirical results show GAN-based methods get the lowest FID yet still fail to recover exact training images, reconstruction quality drops as training size grows, and higher model memorization correlates with easier recovery; LLM-based visual scoring replaces costly human inspection.
AI-written summary · about
chapters (13)
- 0:00Data reconstruction attacks
- 1:15Unclear definitions problem
- 3:04Limitations of evaluation metrics
- 4:09Proposed robust metrics
- 5:34Reconstruction methods comparison
- 6:59Dataset vs sample metrics
- 8:01Memorization and attacks link
- 9:27Human vs automated metrics
- 10:11Evaluating reconstructions with GP4
- 11:14Training data size impact
- 12:16Optimal data size for attacks
- 12:58Pre-training benefits
- 13:21Future directions
related talks
45mSchemes, Privacy, and Security
30mVisual data abundant relevant labelled cheap Pick two
01h 28mPPML: Machine Learning on Data you cannot see
01h 18mComputer Vision after the Victory of Data
20mMachine Learning Attacks – A New Era of Security Epidemic
23mVision-Language Models for Fraud Detection
01h 43mPerspective
40mVisual concept learning from few images