SBIR Phase I: Cloud-Based Platform for Comprehensive AI Robustness Assessment with Dual Optimization for Accuracy and Robustness

About This Grant

The broader/commercial impact of this Small Business Innovation Research (SBIR) Phase I project is the development of a cloud-based platform that improves the robustness of artificial intelligence (AI) models used in critical applications such as smart transportation systems, healthcare, aerospace, and defense. Ensuring AI robustness is essential for the safety, security, and reliability of systems where model failures can cause significant societal harm. This project addresses the need for AI models that perform consistently even under adversarial conditions and input variability. By providing robustness assessment and retraining capabilities, the platform will support the creation of safer and more reliable AI systems. Additionally, the platform is designed to expand beyond robustness evaluation, offering comprehensive AI model health assessments that incorporate generalization, explainability, and security metrics. By establishing standardized quality assurance protocols, this technology has the potential to support the development of global governance and safety standards for AI, ultimately enhancing public trust and promoting the responsible adoption of AI systems across multiple sectors. This Small Business Innovation Research (SBIR) Phase I project addresses critical challenges in developing robust AI (artificial intelligence) models, particularly for safety-critical sectors. The intellectual merit lies in the novel integration of both white-box and black-box robustness evaluation methods into a cloud-based platform designed to assess and improve AI model robustness. A key technical hurdle is balancing model accuracy with robustness, especially in the face of adversarial attacks and input variations. While white-box methods, such as gradient-based evaluations, adversarial attacks, and perturbation analysis, will be used when model internals are accessible, the core innovation of this project is a black-box robustness technique based on manifold curvature estimation. This method evaluates robustness without requiring access to a model’s internal structure, relying solely on input-output relationships. This innovation is crucial for industries where model internals may not be transparent or accessible. The research objectives include developing and validating the platform and implementing a dual optimization method to retrain AI models for both accuracy and robustness, ensuring greater resilience against adversarial attacks. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.