Mitigations and defenses
Now that we have explored the various types of data poisoning attacks, let’s discuss how we can defend against them and mitigate risks. This will include a combination of some traditional defenses integrated into MLOps, as well as adversarial robustness defenses.
Cybercity defenses with MLOps
To a degree, traditional cybersecurity provides defenses that help mitigate data poisoning. Some defenses we saw in Chapter 3 would have made data poisoning harder. These include least-privilege access, encryption, and data hashing or signing. However, we need to see these techniques as part of an integrated system of defenses combining techniques with automated tracking, approvals, monitoring, and alerting.
This is where MLOps can help. Platforms such as AWS SageMaker, MLflow, and Azure Machine Learning offer services and defenses to help us defend against data poisoning. These include the following:
- Data versioning and lineage to track changes...
