AI system validation by fault injection in simulation H/F

Mathematics, information, scientific, software

Internship

AI system validation by fault injection in simulation H/F

The project aims to evaluate an AI-based system design based on safety-oriented fault-injection simulations through a virtual platform to reinforce AI learning during design.

6

To address the issue of AI system verification, validation, and certification, existing work define a design solution combining AI algorithms and conventional algorithms. AI algorithms include all complex and/or black box algorithms that need to be certified. The conventional algorithms include functions that monitor and supervise AI code execution. These monitors are derived from safety and requirement analysis. With this particular combination, we hope to benefit from both approaches: a system that can handle complex decision processes through the use of artificial intelligence and that can provide high confidence by leveraging a monitor written in conventional code. Meanwhile, evaluating the robustness of such AI-oriented systems often involves understanding the impact of fine-grain perturbation regarding the computations. The ability to study high-level and fine grain aspects at the same time requires a significant runtime instrumentation capacity.

To evaluate the safety and robustness of such an AI-oriented system with its HW and SW parts, we think of a methodological approach based on preliminary safety assessment results, functional testing on a virtual platform, fault injection, and uncertainty measurement. Concretely, the evaluation framework will be based on:
• Risk assessment providing some kind of envelope of permissible behavior for the AI system without compromising safety and list of faults and failures with their effects on the system
• An embedded system simulator capable of injecting fault at functional and hardware levels
• Online (at runtime from devices) and offline (from testbench dataset) fault injection (HW/SW) technique to stress the system with regard to faulty scenarios
• uncertainty measures with an estimation of their confidence level
• runtime monitoring for the tracking of fault propagation,
and study of their impact at the algorithmic level

The goal of the project is to implement and validate such an approach on a use case.

Papyrus, Sophia, Simulation platforms

You are passionate about science and novel technologies. You are preparing a Bac+4 or Bac+5 diploma in the field of computer science or alike. You have knowledge of AI, programming languages, embedded systems, simulation, safety assessment.