# [Model Inversion Attack by Integration of Deep Generative Models: Privacy-Sensitive Face Generation From a Face Recognition System](https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9672188)
The authors propose an MIA method integrated with pre-trained deep generative method and utilisesearch domain conecpts for efficient narrowing of search space to face features instead of general image space.This is done in a semi-white box scenario.

### Conventional MIA Problems :

- In convential MIA the image generaed turns out too noisy as the search space is entirety of image space where face features occupy only a small part.

### Deep MIA :
- In Deep MIA a pretrained generative model is used which is trained on faces and then is integrated with the conventional MIA.
- This restricts the search space to the face feature vectors.

**Generative Models :**

- Two different generative models have been studied in the paper DCGAN and α-GAN.
- DCGAN has certain drawbacks which are mode-collapse, low diversity of outputs and no image to feature mapping.
- α-GAN on the other had is a fusion of VAE and GAN and can tackle these problems and hence is preferred.

**Seed Vector :**

- The seed vector for pretraining is important and can be set in two methods: random and face based.
- In face based method average of few images from attackers database are taken and converted to seed vector.
- This performs better as random vector sometimes may generate non-face-like images.

### Experiments:
- The datasets used are VGGFace2,CelebA and PubFig.
- VGGFace2 is used to train target system and PubFig is used for evaluation system.
- Gray-scale images are used in this method for faster training, the procedure for RGB image is the same.
- The metrics used for objective evaluation are top-r accuracy and CCDF( Complementary Cumulative Distribution Function ).
- For subjective evaluation 3 parameters were chosen: Naturalness,Similarity and Recognizability.
- The α-GAN with Face seed vector showed the best performance across all metrics.