Documentation of data collection, annotation, maintenance practices

Data cleaning

Providers can filter out the training dataset via multiple layered techniques, ranging from rule-based filters to anomaly detection via data point influence or statistical anomalies of individual data points [213]. For example, a data cleaning procedure can involve the use of filename checkers to detect duplicates or wrongly formatted data, which then moves to flagging the most influential data samples from the dataset via influence functions for anomaly detectio

Internal data poisoning diagnosis

Providers can have an internal framework to identify what specific data poisoning attack their model may be a victim of based on a set of symptoms, such as analysis of target algorithm and architecture, perturbation scope and dimension, victim model, and data type [39]. This framework includes known defenses against the diagnosed attack, which providers can then apply to the model.

Tamper-resistant safeguards for open-weight models

Training and implementing safeguards can improve the robustness of openweight models against modifications from fine-tuning or other methods to change the learned weights of the models, especially those aimed at removing safety restrictions. These safeguards can be resilient even after extensive fine-tuning, ensuring that the model retains its protective measures [199]