shumaket/adult-census-model-inversion

Model Inversion Attack on Tabular Data (Adult Census)

This repository contains artifacts from a white-box gradient-based model inversion attack on the Adult Census Income dataset.

Method

The attack uses a VAE-learned data manifold prior to ensure reconstructed tabular records stay realistic:

1. Train MLP classifier on Adult Census (target model)
2. Train VAE on same training data (data prior)
3. Optimize latent codes to maximize target model confidence
4. Decode through VAE to reconstruct feature vectors

Results Summary

Metric	Class 0 (<=50K)	Class 1 (>50K)
Mean NN Distance	1.63	1.57
Feature Range Compliance	94.5%	76.1%
Membership Proxy	63.0%	98.5%
Mean Confidence	99.4%	99.96%

Key Finding: The minority class (>50K) is significantly more vulnerable — 98.5% of reconstructed samples fall closer to training data than the typical training sample distance, indicating strong membership leakage.

Files

• reconstructions/class_0_reconstructed.npy — 200 reconstructed samples (class 0)
• reconstructions/class_1_reconstructed.npy — 200 reconstructed samples (class 1)
• reconstructions/class_0_sample.csv — 20 human-readable samples
• reconstructions/class_1_sample.csv — 20 human-readable samples
• metrics.json — Full evaluation metrics
• target_mlp.pt — Trained target model weights
• vae_prior.pt — Trained VAE prior weights

Requirements

• torch, numpy, pandas, scikit-learn, scipy, matplotlib, datasets

References

• Fredrikson et al. (2015). Model Inversion Attacks. ACM CCS.
• PLG-MI (2023). arXiv:2302.09814.

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = 'shumaket/adult-census-model-inversion'
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

For non-causal architectures, replace AutoModelForCausalLM with the appropriate AutoModel class.