The presentation will show how dangerous a poisoning attack can be on a fictitious ADAS system. It will work through a systematic approach to train engineers in detecting and preventing the attack. It will also offer an overview of ISO TR 5469 and discuss other potential functional safety hazards in an AI/ML system.

• Introduction of the 3-stage AI/ML development stages
• Introduce the AI classification scheme from ISO TR 5469
• Explain the six desirable AI properties leading to methods & techniques and acceptance criteria
• A brief look at all the six properties including transparency and explainability and resilience to adversarial inputs