Fuzzing the brain: Automated stress testing for the safety of ML-driven neurostimulation

Key Points

• The aim is to enhance safety evaluations for machine learning models in neurostimulation through a new empirical approach.
• Used violation-focused fuzzing to assess safety of deployed models.
• Transformed safety into a measurable property instead of a training heuristic.
• Established evidence-based benchmarking processes.
• Provided a reproducible method for safety assessments.
• Enabled readiness for regulatory evaluations of neural interfaces.
• Highlighted the importance of empirical safety measures in technology deployment.