Extend Ensemble Adversarial Training by Conventional Input Rectification

[daniel-knape]

• Read paper
• Create the code

[daniel-knape]

21h estimated

[Caryox]

In order to mitigate the adversarial effects, Xu et al. [75] first utilize two squeezing (denoising) methods—bit-reduction and image-blurring—to reduce the degrees of freedom and remove the adversarial perturbations.

We have to check, if bit reduction is suitable because we are using already 8bit Images. Maybe we just use 2 bit for only black/white without grey shades? Or make image squeezing parameters callable to modify image squeezes different for eacht ensemble call.

[Caryox]

• Use data extractor function to manipulate data after data loader is initialized
• Check if image rectification is static or dynamic for each batch (should we always blurr images the same amount or is the blurring randomized?)

[Caryox]

https://arxiv.org/pdf/1705.10686.pdf

Use color depth manipulation and median smoothing

[Caryox]

Testing phase, functionality is given for single network. Next phase is to use image rectification for ensemble