Robustness and Sensitivity Assessment of Deep Neural Networks (DNNs)
Reference persons EDGAR ERNESTO SANCHEZ SANCHEZ
External reference persons Annachiara Ruospo
Description Nowadays, safety-critical systems rely on neural networks-based computations to perform tasks, such as pedestrian detection in the context of autonomous driving. Indeed, Deep Learning is currently one of the most intensively and widely used predictive models for safety-critical applications. Today, ensuring the reliability of these innovations is becoming very important since they involve human beings. The aim of this master thesis is to highlight the bigger DNNs criticisms (layers, neurons..) and to underline the most sensible areas of the DNNs architecture. Then, according to these analyses, the student should propose HW/SW solutions for improving their safety. To fulfil this goal, a Fault Injection-based approach is required.
Required skills - C, Python;
- Programming/scripting skills;
Deadline 16/03/2021 PROPONI LA TUA CANDIDATURA