Fraunhofer secure’s local AI processing for sensors

Author:OMO Release Date: Nov 7, 2019


Secured sensor processing using local artificially intelligence is the aim of a chip announced by Fraunhofer Institute for Microelectronic Circuits and Systems (IMS), which has developed it around an open-source Risc-V core.

“In combination with the AIfES framework for embedded AI we have enabled a system for the application of artificial intelligence on sensor and actuator-related embedded systems”, said IMS’s Alexander Stanitzki.

AIfES (‘artificial intelligence for embedded systems’) is a platform-independent machine learning library for stand-alone microcontrollers – those with no connection to cloud resources or other processing back-up.

AlfES in action “The sensor-related AI system recognises handwriting and gestures, enabling for example gesture control of input when the library is running on a wearable,” according to Fraunhofer IMS.

To protect on-board AI algorithms and learning data from intellectual property theft, the Fraunhofer chip is equipped with memory content protection and hardware-accelerated encryption “for popular encryption methods for the communication”, said the organisation, which is using the term ‘Trusted Electronics’ to describe such hardware.

Chips are scheduled to appear in the spring of next year, and a service to create custom asics along the same lines has been opened – for example, to add customer-specific modules to the Risc-V core.

For evaluation, equivalent soft-cores have been created for installing in FPGAs, and software libraries are available to go with them.

More on AIfES

Python-based AI software is fine for PCs, according to Fraunhofer IMS, but it is not much good for training and executing neural networks on microcontrollers – which is why, specifically for sensor data processing, its researcher team wrote AIfES in C.

Much as MCUs are the target, is will also run on PCs, Raspberry Pis and Android.

The library contains a completely configurable artificial neural network, which can also generate deep networks for deep learning.

“We’ve reduced the source code to a minimum,” said IMS researcher Dr Pierre Gembaczka, “which means the artificial neural network can be trained directly on the microcontroller or the sensor. In addition the source code is universally valid and can be compiled for almost any platform. Because the same algorithms are always used, a neural network generated for example on a PC, can easily be ported to a microcontroller.”

But don’t get carried away, it is not a big data scheme, it is an absolutely necessary data scheme.

“Of course it’s not possible to implement giant deep learning models on an embedded system, so we’re increasing our efforts toward making an elegant feature extraction to reduce input signals”, said IMS embedded systems manager Burkhard Heidemann.

Feature extraction and data pre-processing strategies are included to allow a small neural network to give a useful performance – for example, both learning and instancing for the handwritten number recognition app in the photo above are done on an 8-bit Arduino Uno. Another demonstrator can recognising complex gestures made in the air.

An added benefit of local data processing is that it removes the need to pass that data to a remote server, closing a potential hacking loop-hole.

If data has to be sent for AI processing elsewhere, it is “difficult to protect privacy in this process, and enormous amounts of data are transmitted. That’s why we’ve chosen a different approach and are turning away from machine learning processes in the cloud in favour of machine learning directly in the embedded system. Since no sensitive data leave the system, data protection can be guaranteed and the amounts of data to be transferred are significantly reduced”, said Heidemann.

Fraunhofer

Fraunhofers are German institutions created to help companies bridge the gap between research and mass production. “In eight business units Fraunhofer IMS is dedicated to applied research, advance development for products and their applications,” according to Fraunhofer IMS. “Because of its know-how, access to technology and the high-quality development work, the Institute is a worldwide partner for the industry.”

Photo: The processing core (top left) is based on the Risc-V instruction set, and there is a JTAG TAP (test access port – top right).