MLSEC-10: Protect against data poisoning threats

Use only trusted data sources for training data - Verify that you have sufficient audit controls to replay activity and determine where a change occurred, by whom, and at what time. Before training, validate the quality of training data to look for strong outliers and potentially incorrect labels.

Look for underlying shifts in the patterns and distributions in training data - Using monitoring of data drift, derive the impact to prediction variance. These skews can be an indicator of underlying data drift, and can provide an early warning of unauthorized access targeting the training data.

Identify model updates that negatively impact the results before moving them to production - Determine if the retrained model results are different from the past model iteration. Use past test data and previous model iterations as a baseline.

Have a rollback plan - Using versioned training data and versioned models, make sure you can revert to a known good working model in a failure scenario. Use a fully managed service to store features, such as the Amazon SageMaker Feature Store. See more details of the Amazon SageMaker Feature Store under the Reliability pillar section (MLREL-07).

Use low-entropy classification cases - Look for significant, unexpected changes. Determine the bounds of thresholds, identify classifications that you do not expect to see, and alert if the retrained model exceeds them.

Monitor Bias Drift for models in production

Amazon SageMaker model registry now supports rollback of deployed models

SageMaker Model Registry - Approve

Automated monitoring of your machine learning models with Amazon SageMakerModel Monitor and sending predictions to human review workflows using Amazon A2I

Amazon SageMaker Model Monitor– Fully Managed Automatic Monitoring for Your Machine Learning Models

7 ways to improve security of your machine learning workflows

AWS re:Invent 2020: Detect machine learning (ML) model drift in production

Inawisdom: Machine Learning and Automated Model Retraining with SageMaker

Amazon SageMaker Workshop- DevOps workflow