SEC-Learn - Sensor Edge Cloud for Federated Learning

SEC-Learn - Sensor Edge Cloud for Federated Learning

© Fraunhofer IDMT

Complex information systems and digital technologies are an essential factor for Europe's economic growth and competitiveness. As a result of ongoing digitization, the amount of sensor data to be collected and and evaluated is growing rapidly, which is accompanied by high demands on software and hardware. In addition, increasingly high demands are being placed on data protection and data security. Conventional computing technologies are reaching their limits in terms of speed, performance and energy efficiency. Therefore, Fraunhofer IPMS is pursuing novel approaches from neuromorphic computing in the SEC-Learn project in order to for example, to realize computationally intensive tasks in applications such as mobile devices or vehicles, where energy and storage capacities are limited.

Many applications that require the evaluation, categorization or display of huge amounts of data nowadays rely on artificial intelligence (AI). On the one hand, the high volumes of data generated as a result are accompanied by increased energy requirements, and on the other hand, the data is stored centrally, which entails risks from the point of view of data protection and data security. New approaches that combine both energy efficiency and security aspects are therefore desirable. In the Sec-Learn project, Fraunhofer IPMS is conducting research with ten other Fraunhofer institutes to develop a neuromorphic computing architecture for federated learning. Federated learning refers to the approach in which AI algorithms are trained to store data in a distributed manner on multiple devices or servers. Unlike traditional cloud-based machine learning, this offers the advantage of keeping sensitive data in local systems - a big win for data privacy and security. The platform developed in the SEC-Learn project will also use neuromorphic hardware accelerators that have orders of magnitude lower power consumption. The developments are being tested on two use cases - speech recognition for voice assistants and image recognition for autonomous driving.

In the project, Fraunhofer IPMS is leveraging its expertise in novel memory technologies, in particular embedded non-volatile memories (NVM) in advanced technology nodes. These enable local integration of logic elements and thus in-memory computing.