Censoring chemical data to mitigate dual use risk

Area of Science:

• Chemistry
• Computer Science
• Data Science

Background:

• Open-source machine learning models in chemistry present dual-use concerns, particularly regarding toxicological data and chemical warfare agents.
• Existing risk frameworks often lack specific data-level mitigation strategies for these concerns.

Purpose of the Study:

• To introduce and evaluate a novel data-level mitigation strategy for dual-use machine learning models in chemistry.
• To assess the effectiveness of data noising compared to data omission for preventing malicious use.

Main Methods:

• Development of a selective data noising technique to introduce prediction errors in sensitive data regions.
• Evaluation of the noising method on molecular feature multilayer perceptrons and graph neural networks.
• Comparison of noising strategy against simple data omission for preventing model misuse.

Main Results:

• Selective data noising effectively increases prediction error and variance in targeted sensitive regions.
• Data omission alone is insufficient to prevent model extrapolation and potential misuse.
• The noising method demonstrates efficacy across different molecular machine learning architectures.

Conclusions:

• Data noising is a viable strategy for mitigating dual-use risks associated with chemical machine learning models.
• This approach enhances the safety of sharing potentially sensitive molecular data.
• Further research into data-level defenses is crucial for responsible AI development in chemistry.