sAIfer Lab research enhances desired properties of ML systems, including fairness, transparency, and robustness.

sAIfer Lab's research approach fits with the so called "Pasteur’s quadrant", where practical considerations drive research questions requiring novel and fundamental understanding. This approach combines the pursuit of fundamental scientific understanding with practical societal applications in cybersecurity, biometrics, and industry. The Lab focuses on developing safe and secure AI systems designed to prevent both unintentional harm, which can arise from AI systems' inadequate handling of unexpected situations, and intentional harm, stemming from AI vulnerabilities to attacks. This ensures the protection of people, IT systems, society, and critical infrastructures.

Since "there is nothing more practical than a good theory" (Kurt Lewin, psychologist, 1950), both Pra Lab and SmartLab have have invested great efforts in studying theoretical foundations to develop  safer and more secure AI-based and ML-based systems.

The research activities conducted by SmartLab since 2005 encompass a range of principles and methodologies that provide a deep understanding of how learning algorithms work. Central to these foundations is the Statistical Learning Theory, which offers a framework for understanding and analyzing the properties and behavior of learning algorithms, including, but not limited to, fairness, which is their ability to provide equitable decisions across different demographic groups, and transparency, which is their ability to provide decisions understandable by the users. 

This deep comprehension of how Machine Learning works has also been leveraged by the Pra Lab since 2007 to provide a pioneering contribution to the field of Adversarial Machine Learning, which analyzes the security of Machine Learning in the presence of attackers aimed at subverting its functionality for their illegitimate purposes. For example, to make an ML-based malware detector unable to recognize its malware or to extract from a ML system private information about its users.

Together, these theoretical foundations form a comprehensive framework that guides the development of more effective, ethical, and reliable machine learning systems.