Maskininlärning i enheter för ett säkrare sakernas internet
- Diarienummer
- FID22-0039
- Projektledare
- Rahimian, Fatemeh
- Start- och slutdatum
- 230801-280731
- Beviljat belopp
- 2 500 000 kr
- Förvaltande organisation
- RISE Research Institutes of Sweden AB, Borås
- Forskningsområde
- Informations-, kommunikations- och systemteknik
Summary
Advances in machine learning have enabled progress across many application domains. Typically models are trained and deployed on large cloud servers, but recently some models are also deployed on less resource-rich devices, such as smartphones. This project enables on-device learning for IoT devices that are even more limited in terms of resources. This has many advantages. First of all, on-device learning is privacy-preserving by design, and thus, many applications areas that require strict data privacy could benefit from it. Secondly, it does not require an internet connection and there is no need to upload data to the cloud or download a model, saving bandwidth, latency and energy. This is crucial for time-critical applications, as well as low-power IoT devices, for which communication is very expensive. Lastly, the models can adapt to the target environment and provide personalized or special-purpose services. Hence, this project aims to make resource-constrained IoT devices smarter by enabling on-device learning. This includes the initial training based on offline data, model compression, transfer of the model to the IoT network, and continued on-device training for adaptation to the environment. We will implement a complete end-to-end system in a real application scenario related to IoT security. In the long run, we expect a wide industrial uptake for such secure and smart IoT solutions that help to keep Sweden at the forefront of the IoT industry.
Populärvetenskaplig beskrivning
Machine learning gives computers the capability to imitate intelligent human behavior, typically by learning from a large number of examples, so-called training data. Advances in machine learning have enabled progress across many application domains. Typically models are trained and deployed on large cloud servers since both the training and the execution of the models require a lot of computing power that is available in the cloud. Recently some models are also deployed on less resource-rich devices, such as smartphones that have less computing power and memory, which makes it more difficult to handle large amounts of data. This project enables on-device learning for Internet of Things devices that are even more limited in terms of resources. Typical examples include temperature and humidity sensors. This has many advantages. First of all, on-device learning is privacy-preserving by design, and thus, many applications areas that require strict data privacy could benefit from it since all data is handled on the device itself. Secondly, it does not require an internet connection and there is no need to upload data to the cloud or download a model which saves bandwidth, latency and energy. This is crucial for time-critical applications, as well as low-power IoT devices, for which communication requires a lot of energy and thus is regarded as expensive. Lastly, the models can adapt to the target environment and provide personalized or special-purpose services. For all the above reasons, this project aims to make resource-constrained IoT devices smarter by enabling on-device learning. This includes the initial training based on offline data, model compression, transfer of the model to the IoT network, and continued on-device training for adaptation to the environment. We will implement a complete end-to-end system in a real application scenario related to IoT security. In the long run, we expect a wide industrial uptake for such secure and smart IoT solutions that help to keep Sweden at the forefront of the IoT industry.