Paint: Accurate, efficient, and sparse Model Inversion via peripheral sparsification and anchor alignment
How can we reconstruct training data from a pre-trained model without access to the original dataset? Model Inversion (MI) attempts to synthesize samples similar to the training dataset. This capability is particularly useful in data-free applications, such as data-free quantization and data-free knowledge transfer, where synthetic data replace inaccessible training samples. For these applications, MI needs to generate high-fidelity samples to preserve semantic information while remaining computationally efficient, enabling scalable generation for reliable downstream performance. However, existing MI methods fail to achieve both high fidelity and efficiency simultaneously. In this paper, we propose Paint (Peripheral sparsification and Anchor-aligned INversion Technique), an accurate, efficient, and sparse MI framework for generating high-fidelity synthetic data. Paint integrates low-pass filtering during inversion to suppress high-frequency noise, aligns salient regions toward anchor with peripheral sparsification to preserve object structures during sparse inversion, and exploits soft targets to better capture the predictive uncertainty. Extensive experiments demonstrate that Paint improves inversion efficiency for ViTs and achieves accuracy gains of up to 5.33%p in data-free scenarios.
Authors
- Jaehyeon Choi (ORCID: https://orcid.org/0000-0002-7214-5554)
- U Kang (ORCID: https://orcid.org/0000-0002-8774-6950)
- Minjun Kim (ORCID: https://orcid.org/0009-0007-1119-5615)
Institutions
- Seoul National University (KR)
Publication Details
- Journal
- PLoS ONE
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1371/journal.pone.0362597
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Seoul National University
- Ministry of Science and ICT, South Korea
- Institute for Information and Communications Technology Promotion
- Institute of Engineering Research, Seoul National University