Deep neural networks (DNNs) are artificial neural networks (ANN) and a key component of artificial intelligence (AI) applications.  They are designed to function similarly to the human brain and trained using large labeled data sets. There is a plethora of applications for DNNs, including but not limited to driverless cars, security, facial/voice recognition, healthcare, and finance. One of the key issues with current DNN technologies is malicious attacks that attempt to mislead applications with deceptive data. As a result, there is a need for new technologies that can circumvent these attacks and prevent the dangerous consequences in critical applications.

This invention describes the methods, systems, and algorithms for optimizing the architecture of DNNs for improving performance and preventing adversarial attacks. It utilizes graphs of geometrical robustness measures, specifically Ollivier-Ricci and Forman-Ricci curvatures, to prevent attacks against DNNs and optimize performance. The technology can be integrated into many existing AI/ML applications resulting in improved performance and protection against malicious attacks.

Competitive Advantages

• Protects DNNs against adversarial attacks.
• Enhances the performance of artificial applications.
• Applicable to many industries that utilize DNNs for data processing.

Opportunity

• The market for artificial intelligence was $3.5 billion in 2018, expected to exceed $26 billion by 2023 at a CAGR of 49%.
• Deep learning technologies experienced revenues of $384 million in 2018 and are anticipated to reach $3.1 billion by 2023 at a CAGR of 51%.

Rowan University is seeking a partner(s) for further development and potential commercialization of this technology. The inventor is available to collaborate with interested companies.