Challenges and efforts in managing AI trustworthiness risks: a state of knowledge

Area of Science:

• Artificial Intelligence (AI) Risk Management
• Cybersecurity
• Socio-technical Systems

Background:

• Existing AI risk management frameworks neglect human factors and lack metrics for social/human threats.
• NIST AI RFM and ENISA insights reveal limitations in human-AI interaction and the need for ethical/social measurements.
• AI trustworthiness is multifaceted, encompassing legislation, cyber threat intelligence, and adversary characteristics.

Purpose of the Study:

• To identify critical gaps in AI risk management frameworks, particularly concerning human and social dimensions.
• To propose a comprehensive approach to AI trustworthiness by integrating technical and social mitigation strategies.
• To introduce innovative defense mechanisms for enhancing AI security and understanding adversary profiles.

Main Methods:

• Analysis of existing AI risk management frameworks (NIST AI RFM, ENISA).
• Exploration of technical threats (data access, poisoning, backdoors) and socio-psychological threats (bias, misinformation, privacy).
• Development of a combined technical and social mitigation approach, including cyber-social exercises and digital clones.

Main Results:

• Identified significant neglect of human factors and social threats in current AI risk management.
• Demonstrated the necessity of interdisciplinary collaboration among cybersecurity, AI, and social science professionals.
• Proposed novel defense strategies like cyber-social exercises and digital clones for enhanced AI security.

Conclusions:

• A robust AI ecosystem requires integrating technical and social mitigation measures, standards, and continuous research.
• Interdisciplinary collaboration, awareness campaigns, and ongoing research are crucial for ethical AI development.
• Developing resilient AI systems necessitates addressing both technical vulnerabilities and societal/human impacts.