Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Generative Model Inversion Through the Lens of the Manifold Hypothesis

Authors: Xiong Peng, Bo Han, Fengfei Yu, Tongliang Liu, Feng Liu, Mingyuan Zhou

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we first validate the hypothesis proposed in Sec. 4, followed by a comprehensive evaluation of the training-free Align MI approach introduced in Sec. 5. Our experiments focus on real-world face recognition tasks. To ensure computational efficiency, we perform hypothesis validation in the low-resolution setting (64 64), where tangent space estimation is tractable. For the method evaluation, we compare the performance of state-of-the-art generative MIAs before and after integrating our proposed techniques, i.e., PAA and TAA.
Researcher Affiliation Academia 1TMLR Group, Department of Computer Science, Hong Kong Baptist University 2Sydney AI Centre, The University of Sydney 3School of Computing and Information Systems, The University of Melbourne 4Mc Combs School of Business, The University of Texas at Austin
Pseudocode Yes Algorithm 1 Gradient Manifold Alignment-Aware Training Algorithm 2 Training-Free Gradient Manifold Alignment During Inversion
Open Source Code Yes The code is publicly available at https://github.com/tmlr-group/Align MI.
Open Datasets Yes In line with existing MIA literature, we use the Celeb A [Liu et al., 2015], Face Scrub [Ng and Winkler, 2014], and FFHQ datasets [Karras et al., 2019].
Dataset Splits Yes These datasets are divided into two parts: the private training dataset Dpri and the public auxiliary dataset Daux, with no overlapping classes.
Hardware Specification Yes All high-resolution MIA experiments using Plug & Play Attacks (PPA) were conducted on Oracle Linux Server 8.9 with NVIDIA A100-80G GPUs... Low-resolution facial recognition MIAs were run on Ubuntu 20.04.4 LTS with NVIDIA RTX 3090 GPUs
Software Dependencies Yes using CUDA 11.7, Python 3.9.18, and Py Torch 1.13.1. Low-resolution facial recognition MIAs were run on Ubuntu 20.04.4 LTS with NVIDIA RTX 3090 GPUs, under CUDA 11.6, Python 3.7.12, and Py Torch 1.13.1.
Experiment Setup Yes Models are optimized using Adam [Kingma and Ba, 2015] with an initial learning rate of 10-3, β parameters set to (0.9, 0.999), and a weight decay of 10-3. Training runs for 100 epochs with a batch size of 128, and the learning rate is reduced by a factor of 0.1 at epochs 75 and 90.