Markdown to PDF

New

Convert Markdown files to PDF

Your files are processed locally in your browser and never uploaded to any server.

Drop a .md / .txt file here, or click to browse

Supports Markdown, plain text

Markdown Input

# Machine Learning Privacy

## Introduction

This document covers **model inversion attacks** using the *IBM ART framework*.

### Key Parameters

The algorithm uses these main parameters:

- **label** — the class to reconstruct
- **α** — maximum iterations
- **β** — patience parameter
- **γ** — minimum cost threshold

### Workflow

1. Understand the basic model inversion attack
2. Modify the CNN architecture
3. Observe how outputs change

### Cost Function

The cost function is written as:

`c(x) = 1 - f_label(x) + AUXTERM(x)`

### Code Example

```python
import numpy as np
from art.attacks.inference.model_inversion import MIFace

# Initialize attack
attack = MIFace(classifier, max_iter=10000)
reconstructed = attack.infer(x_init, y=target_label)
```

### Comparison Table

| Parameter | Reference CNN | Modified CNN |
|-----------|--------------|--------------|
| Layers    | 3 conv + 1 FC | 5 conv + 2 FC |
| Accuracy  | 98.7%        | 99.1%        |
| Parameters | 26K         | 84K          |

> **Note:** The trained model may store information that leads to privacy leakage.

---

Results show that modifying the CNN architecture significantly affects the reconstructed representations.

Preview

Page Size

Font Size: 15px

Page Margin: 48px