Summary
We engaged in our first in-depth exploration of an essential part of adversarial AI, poisoning attacks that affect model development and can be hard to detect. We covered basic concepts and examples. We also detailed the implementation of simple and advanced poisoning attacks. This knowledge can help us test and evaluate our models for poisoning. We also learned about other defenses against data poisoning, including MLOps, anomaly detection, and advanced defenses offered by ART.
Poisoning attacks assume access to the training data and rely on interfering with model training to undermine the model’s integrity. In the next chapter, we will look at a different poison-less approach to implanting backdoors and attacking the model’s integrity by tampering with the model and injecting Trojan horses or performing model reprogramming.
