# [Attacks on state-of-the-art face recognition using attentional adversarial attack generative network](https://link.springer.com/content/pdf/10.1007/s11042-020-09604-z.pdf)
The authors propose a generative model for adversarial attacks using attention mechanism.

### Attentional Adversarial Attack generative network :

- The model has two discriminators,sample and identity.
- To make the generated sample close to target an attentional VAE is used as a feature extractor.

**Black box:**
- In a white box scenario the given target system acts as identity discriminator, in a black box scenario a substitute network is used.
- Feature estimation is very important in a black box scenario which is done by the substitue model which uses estimation loss introduced by the authors.

### Experiments :
- The metrics used are Real/Fake accuracy, mAP, similarity score and SSIM.
- The datasets used are CASIAWebface,MS-Celeb-1M,LFW,CFP-FP and AgeDB-30.
- The evalation is done in modular maner checking the importance of each block.
- When compared with previous works on CASIA false label A³GN shows best success rate.