The objective of this research is to increase the predictability and the trustworthiness of AI algorithms and deep networks enable their use in safety-critical cyber-physical systems, as autonomous vehicles, advanced robots, space crafts, and medical systems. To be safely deployed, such systems must be certified and must react within given timing constraints imposed by the environment. Unfortunately, current deep learning frameworks are not designed to be used in safety-critical systems and cannot guarantee predictable response times.
The RETIS lab also gained considerable experience in enhancing the safety and security of autonomous systems that leverage artificial intelligence and deep neural networks in perception and control tasks, by investigating attack and defense methods for adversarial examples and architecture frameworks to increase the trustworthiness of deep neural networks and tolerate faults in AI components.
The most relevant investigated topics are the following:
- Safe and secure architectures for AI-powered cyber-physical systems;
- Defense perturbations to detect adversarial examples;
- Coverage analysis for increasing trustworthiness of deep neural networks;
- Predictable support for concurrent deep neural networks on GPU platforms;
- Predictable FPGA acceleration of deep neural networks;
- Lidar odometry and localization through deep learning;
- Explainability of deep neural networks;
- Verification of deep neural networks;
- Accident prediction in autonomous driving;
- Enhance predictability in inference engines;
- AI for cloud computing and network function virtualization (NFV) infrastructures;
- Improving predictability, safety, and security in the Apollo autonomous driving framework.
Key Recent Projects
- On-going: Enhanced Train localization by sensor fusion and machine learning. Industrial project funded by Hitachi Rail STS, Industrial project, 2020-2022.
- On-going: AI e Machine Learning methods for supporting performance monitoring, capacity planning, anomaly detection and troubleshooting in infrastructures for Network Function Virtualization”, Industrial project, 2021-2022.
- Concluded: Predictable, Safe, and Secure Software Systems for Autonomous Driving. Industrial project, 2019-2021.