Last October, Infineon & XAIN announced an R&D partnership with Infineon Technologies. In a first MVP, we demonstrated how the features of Infineon’s powerful AURIX microcontrollers can be leveraged to program and enact usage control of physical functions and fine-grained access control to data streams; the video at https://www.youtube.com/watch?v=-RlHTRy-HMs&t=1s shows the outcomes of this MVP.
Microcontrollers can run AI applications
Infineon’s microcontroller family AURIX comes with crypto accelerators, multi-core architecture and excellent data and program storage capabilities. These advances in embedded systems have been complemented by recent innovations in signal processing for AI that make it feasible to run AI applications on AURIX microcontrollers. This means that the automotive industry and its stakeholders — including insurance companies, suppliers, drivers of cars, and service providers of infrastructures — have an incredible opportunity to intelligently tap into the rich data streams of modern vehicles.
Integrating AI products into the embedded systems of vehicles reduces considerably the need for costly Vehicle to Cloud communication of data streams, and it can help with addressing data-privacy issues by keeping personal data confined within the vehicle at which it originates. It also makes it easier, and more transparent, to control the usage of such data: when data stays within the vehicle, one can control access to it and its lifecycle right where the data originates and where it is stored locally. Our first MVP explored these benefits by integrating such data usage control with Infineon’s AURIXTM TC399 development board.
Adoption barriers for AI applications in automotive
Advanced signals processing and AI capabilities of powerful AURIX microcontrollers are certainly effective enablers of AI in Vehicle to X infrastructures and their service provisions, as just discussed. But many commercial use cases for this require additional enabling technology for the effective removal of AI adoption barriers.
A key such barrier is posed by regulation, in particular the EU’s data privacy regulation GDPR. The GDPR is bound to shape privacy technology in any products that deal with personal data of EU citizens. The Cyber Security Law of The People’s Republic of China and the California Consumer Privacy Act of The State of California are two more examples of how law and regulation will demand privacy safeguards, which will often be relevant to AI products used within vehicles.
The relevance of privacy compliance originates from the fact that a lot of data generated by a vehicle and its internal network is to be considered personal data: timed location data, driving behavior, data on infotainment usage, and so forth. Running AI applications within a vehicle therefore can address or reduce the complexity of some ensuing data-privacy concerns. However, a major disadvantage of training an AI model from data confined to a sole vehicle is that the machine learning then works with “small data”, for example driving behaviour of a sole user. This will then impair the quality of the AI model so that it may be too inaccurate to be used in a product, for example for insurance premiums that adjust according to driving behavior. Clearly, a model that is trained on the behavioural data of many users will be much superior. And when datasets from all over the world are used for training an AI model, such a model could also account for regional factors when supporting decisions for setting insurance premiums.
Federated Machine Learning overcomes these adoption barriers
Conventional machine learning and its distributed versions require that data from vehicles be moved elsewhere, say, to some cloud before machine learning can take place. Such externally stored and aggregated data could then be processed in a usual machine-learning workflow. But this is hardly possible when the data used from vehicles contains personal data:
- aggregation of data from different vehicles may require highly complex contractual agreements about data usage and control
- alternatives such as anonymization of datasets are costly, time consuming, do not scale, and may not be GDPR-compliant.
XAIN’s FedML approach and vision of its use in automotive
XAIN is addressing this problem by developing its unique approach to Federated Machine Learning (FedML). This machine-learning technology allows learning to take place at local datasets. Local data will never have to be moved elsewhere. Rather, the models are trained locally on local data. These locally updated models are then communicated to a Coordinator that aggregates them into an updated global model. All participants of this federated learning then receive this updated global model as the basis for their next local training round.
Our vision for FedML in automotive is therefore to equip powerful microcontrollers such as the AURIXTM with our software modules that take care of the local training with local data as well as the communication with a Coordinator. This solves the above problems around data privacy of AI applications. At the same time, this offers an accuracy of AI models that is basically as good as the one obtained by standard machine learning done on an aggregation of all data from all participating vehicles — which would not be GDPR compliant.
A particularly exciting range of use cases we foresee for FedML in automotive are found in autonomous driving. Driving vehicles at the higher levels of autonomy, including the SAE level 6 at which no human intervention is required anymore, will require AI models that have the highest degree of accuracy and reliability. Given the above regulatory constraints, FedML can act here as a genuine enabler by helping to achieve such high levels of accuracy for AI models that support decision making in autonomous driving.
Furthermore, the upcoming 5G high-speed connectivity is opening up significant opportunities for innovative FedML use cases in the next generation of vehicles, cooperative intelligent transport systems, mobility as a service, and transportation as a service.
For this, FedML not only solves data privacy issues; it also enables competitors such as a consortium of car manufacturers to collaborate in use cases, since FedML can protect not only personal data but also sensitive data: while locally updated models may contain personal or sensitive data, XAIN’s approach to FedML produces globally updated models that, from a legal perspective, no longer contain any personal or sensitive data. In particular, XAIN’s global models are not subject to the GDPR.
Infineon & XAIN can combine their capabilities to produce mature technology that will be first-in-class for enabling GDPR-compliant AI applications in the nascent ecosystem for connected and autonomous vehicle embedded systems.
We are convinced that our collaboration with Infineon will be a big step towards XAIN’s goal of building a trustworthy, compliant, and scalable machine-learning network that is suitable for commercial use and production ready. XAIN believes that attaining this goal will bring crucial first-to-market capabilities for XAIN as well as for Infineon, for a wide and varied range of business cases.