How to train your model dynamically using adversarial data

What you will learn here

• 💡the basic idea of dynamic adversarial data collection and why it is important.
• ⚒ how to collect adversarial data dynamically and train your model on them - using an MNIST handwritten digit recognition task as an example.

Dynamic adversarial data collection (DADC)

Static benchmarks, while being a widely-used way to evaluate your model's performance, are fraught with many issues: they saturate, have biases or loopholes, and often lead researchers to chase increment in metrics instead of building trustworthy models that can be used by humans ¹.

Dynamic adversarial data collection (DADC) holds great promise as an approach to mitigate some of the issues of static benchmarks. In DADC, humans create examples to fool state-of-the-art (SOTA) models. This process offers two benefits:

1. it allows users to gauge how robust their models really are;
2. it yields data that may be used to further train even stronger models.

This process of fooling and training the model on the adversarially collected data is repeated over multiple rounds leading to a more robust model that is aligned with humans¹ .

Training your model dynamically using adversarial data

Here I will walk you through dynamically collecting adversarial data from users and training your model on them - using the MNIST handwritten digit recognition task.

In the MNIST handwritten digit recognition task, the model is trained to predict the number given a 28x28 grayscale image input of the handwritten digit (see examples in the figure below). The numbers range from 0 to 9.

Image source: mnist | Tensorflow Datasets

This walkthrough will be divided into the following sections:

1. Configuring your model
2. Interacting with your model
3. Flagging your model
4. Putting it all together

Configuring your model

Now that you have defined the structure of your model, you need to train it on the standard MNIST train/dev dataset.