Artificial Intelligence Safety (RESCHIP4EU)

This module developed by Politecnico di Torino, provides a comprehensive introduction to the dependability and safety of artificial intelligence (AI) systems, with a special focus on their application in safety-critical domains. Students will explore the foundational principles of AI, including deep learning and the hardware architectures that support modern algorithms. The module covers the evolving landscape of AI standardization and industry regulations, such as the EU AI Act. Then, students will learn state-of-the-art solutions to assess, detect, and mitigate hardware-induced faults in AI systems. The course concludes with a discussion of future trends and challenges in the AI safety field.

About this training module

As AI systems become increasingly integrated into safety-critical domains - such as healthcare, transportation, and industrial automation - their dependability and safety have emerged as urgent priorities. Unlike traditional deterministic software, AI-driven systems introduce new complexities and unpredictable behaviours, making it challenging to guarantee their reliability and trustworthiness in scenarios where failures can have catastrophic consequences. There is a growing need for engineers and practitioners who not only understand the foundational principles of AI and deep learning but are also equipped to address the unique safety and dependability challenges these technologies present. This includes navigating the evolving landscape of AI standardisation and regulation, such as the EU AI Act, and mastering state-of-the-art techniques for detecting and mitigating hardware-induced faults in AI systems.

Module structure

• Introduction and Motivation 
• Hardware architectures for AI 
• European AI Act 
• AI Safety and Reliability 
• Vulnerability Setup 
• Error and Fault models in AI systems 
• Safety and Reliability Assessments 
• Uncertainty in DNNs 
• Safety and Reliability Improvement 
• Conclusions and Future Trends

Learning outcomes

Upon successful completion of this training module, learners will be able to: 

• Describe the fundamental concepts of artificial intelligence, including its economic and social impacts, and recognise both its benefits and associated risks, especially in safety-critical systems. 
• Explain the principles of dependability in complex systems and the role of deep learning in modern AI applications. 
• Identify and compare the main hardware architectures used to run AI algorithms, such as GPUs, ASICs, TPUs, and hardware accelerators. 
• Summarise the current AI standardisation landscape and outline key international industry standards relevant to AI system development and deployment. 
• Evaluate academic and industrial methodologies for assessing the safety of AI systems, with a focus on fault injection (FI)-based approaches at software, architectural, and physical levels. 
• Distinguish between fault mitigation and detection techniques for enhancing the safety of AI systems, including both active and passive approaches and the use of specialised test libraries. 
• Discuss emerging trends and future challenges in the field of AI safety, informed by the latest research and regulatory developments (e.g., the EU AI Act). 

Further details 

The course is developed within the framework of the RESCHIP4EU project, supported by the Digital Europe Programme of the European Union. RESCHIP4EU aims to support the excellence of EU higher education around embedded systems design in a holistic way, from silicon via System-on-Chip design and manufacturing to smart and safety-critical platform and application software. 

More information about the training module on Artificial Intelligence Safety is available on this page.