Federated learning enables machine learning on distributed datasets without needing the learner to access directly the datasets owned by respective stakeholders. The Internet of Things (IoT) provides a fertile ground for applying federated learning, where distributed IoT devices produce a plethora of data that are often private. However, IoT devices are vulnerable to environments with inaccurate data samples and malicious attacks, which is a significant challenge for federated learning. Agglomerating data in a federated and robust manner may produce benefits to the economy and society.
Objectives of the Robust Federated Learning for Internet of Things (FLINT) project include: (1) Formulate federated learning (FL) in heterogeneous, dynamic IoT environments with unreliable and adversarial clients. (2) Design new FL algorithms that are robust against hostile conditions with benign, unreliable, and malicious clients injecting erroneous or poisonous data. (3) Design novel incentive mechanisms to ensure rational clients gain non-negative utility by contributing training data and resources. (4) Analyze complexity, performance, and theoretical bounds of proposed algorithms. (5) Build an IoT testbed to study the learning performance of robust FL solutions. (6) Simulation experiments on real-world datasets to evaluate performance scalability.
The FLINT project will offer graduate and undergrad students a unique opportunity to gain interdisciplinary education in the design of robust FL algorithms for IoT. Research findings will enrich courses on cyber-physical systems and machine learning for IoT. The project will attract women and underrepresented minority students, and attract primary and secondary school students in Science, Technology, Engineering and Mathematics (STEM) disciplines. Dissemination of research results will be by means of the project website, seminars and keynote talks, conference presentations, and publications in top-tier journals and conference proceedings.
The FLINT project website (https://tluocs.github.io/FLINT/) will maintain computational codes, models, real world and experimental data for two years after the project period is over. The website will provide two levels of access â€“ Public and Login required â€“ and an account can be created via registration with no charge. An accompanying GitHub repository will make the developed codes and simulation models available to the community. Permission will be granted to use and distribute freely the data and code with due acknowledgement of the copyright notice and the authors.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.