File size: 1,275 Bytes
c731a37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
---
pretty_name: ImpMIA Data
tags:
- membership-inference
- privacy
- image-classification
- cifar10
- cifar100
- cinic10
- arxiv:2510.10625
license: other
---

# ImpMIA Data

Data for [ImpMIA: Leveraging Implicit Bias for Membership Inference Attack](https://arxiv.org/abs/2510.10625).

## Layout

```text
data/
  no_auxiliary_knowledge/{cifar10,cifar100,cinic10}/
  full_auxiliary_knowledge/cinic10/
```

Each directory contains `x_train.npy`, `y_train.npy`, `x_test.npy`, and `y_test.npy`. Images are `uint8` arrays in NHWC format; labels are `int64`.

The No-Auxiliary-Knowledge candidate pools combine 30,000 target-distribution images with auxiliary images from CINIC-10, Open Images, or generated data. Their first 30,000 rows form the target-training source pool. The Full-Auxiliary-Knowledge CINIC-10 pool contains 50,000 CINIC-10 images.

## Licensing

See `LICENSES.md` and the Open Images attribution tables under `metadata/`.

## Citation

```bibtex
@article{golbari2026impmia,
  title   = {ImpMIA: Leveraging Implicit Bias for Membership Inference Attack},
  author  = {Golbari, Yuval and Wasserman, Navve and Vardi, Gal and Irani, Michal},
  journal = {arXiv preprint arXiv:2510.10625},
  year    = {2026},
  url     = {https://arxiv.org/abs/2510.10625}
}
```