Federated Learning on Edge

Revolutionizing privacy and efficiency in data processing with Federated Learning (FL) on edge devices, enabling real-time, secure analytics in several key sectors.

Automotive OEM

2023

Client and Industry

A premier German OEM, recognized as a market leader that prioritizes quality and epitomizes luxury and innovation in the automotive sector. Known for its relentless pursuit of excellence, the brand has consistently introduced pioneering technologies and luxurious vehicles, setting the gold standard for safety, performance, and design.

Industry

Automotive OEM

Business Type

Enterprise

Services

Research, Development & Deployment

Our Role

We implemented Federated Learning (FL) on edge devices with limited computing capabilities by deploying efficient, lightweight local machine learning models. These models process data on-device, enhancing privacy and security, while periodically updating a central global model. This approach allows for comprehensive learning without exposing raw data, optimizing performance and longevity.

Project Goals

01

Model Validation: Use a public dataset to validate the model's efficacy.

02

Result Reproduction: Reproduce the results within the client's specific environment using their proprietary data.

03

Model Integration and Deployment: Integrate and deploy the model on test vehicles.

Business
Challenges

01.

Data Privacy Concerns: Due to GDPR restrictions, sensitive data, which can be personally identifiable and could influence insurance premiums, cannot be sent to servers. This necessitates on-device training to ensure privacy.

02.

Limited Computing Resources: The project involves using micro-controllers with constrained computational memory, requiring the development of very small neural networks with fewer than 100 trainable parameters.

03.

Software Limitations: Neural network code must be written from scratch in C, as these micro-controllers do not support many C packages.

04.

Code Safety and Compliance: Extensive code validation, and compliance with various standards like MISRA (embedded coding in automotives), SPICE (software), and ASIL(automotive), are essential to prevent system lock-ups and ensure safety

Solution

We developed a customized solution employing Federated Learning that operates within the stringent limitations of automotive edge devices.

Key Features

01

The local models were crafted to be extremely lightweight yet effective, capable of running on micro-controllers with minimal computational power.

02

By employing a novel neural network architecture and optimization techniques, these models can train directly on devices without the need to transmit sensitive data, maintaining compliance with GDPR.

03

The models are rigorously tested and refined to meet automotive industry standards for safety and reliability.

Impact

01.

Our innovative approach effectively balances the computational constraints of edge devices with the need for sophisticated, privacy-preserving model training. This extends the applicability of Federated Learning to resource-limited and privacy-sensitive environments.

02.

The successful implementation of this technology in predicting parameters not only enhances vehicle performance and longevity but also sets a new standard for deploying advanced machine learning techniques in the automotive sector.