Staging a simple poisoning attack

In Chapter 2, we developed our sample ImRecS AI system by training a CNN with CIFAR-10. The most straightforward poisoning attack strategy is to insert misclassified samples into the training set to degrade the model’s performance.

Note

This is a white box data poisoning attack, whereby the attacker has access to our data, models, and pipelines either as part of a system breach and lateral movement or as a malicious compromised insider.

Here is a reminder of how we generated the training dataset:

from keras.datasets import cifar10
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
# Split the remaining data to train and validation
x_train, x_val, y_train, y_val = train_test_split(x_train, y_train, test_size=0.20, shuffle=True)
cifar10_class_names = ["airplane", "automobile", "bird", "cat", "deer", "dog", "frog", "horse", "ship", &quot...