Privacy Restricted Machine Learning for Edge Computing

Project Objectives

The A2-PRIVCOMP project focuses on defining and developing technological solutions that contribute to the transversal areas of privacy, security and data management in the cloud-edge ecosystem. The intended impact in the medium and long term consists of increasing the trust of users and companies in the services, applications and platforms that run their services on the cloud-edge infrastructure.

The project proposes four specific technological contributions.

• Use of Machine Learning to detect and mitigate the risks associated with the distribution and management of malware in the public cloud-edge ecosystem.
• Development of a platform for data governance with privacy guarantees for cloud-edge services.
• Synthetic data generation models for machine learning solutions with privacy guarantees in the edge-cloud environment.
• Methodology to measure the exposure of personal data in popular applications in the cloud-edge ecosystem.

At Vicomtech, we are researching on the use of synthetic data generation models for machine learning solutions with privacy guarantees with the following objectives:

• Identify and analyze synthetic data generating models for machine learning solutions with existing privacy guarantees.
• Launch and implement synthetic data generating models for machine learning solutions with privacy guarantee.
• Design a methodology to evaluate the synthetic data generated by the generating models, ensuring a balance between usefulness and privacy of the data.
• Provide a validated prototype applicable to cloud-edge infrastructures for the secure and private sharing of synthetic data for industry and public services use cases.
• Generate scientific content to strengthen the EU digital industry within a framework of cybersecurity and resilience through attendance at conferences and publications in scientific journals.

Work done, conclusions and next steps

To date, the "State-of-the-art study on synthetic data generating models for machine learning solutions" has been carried out, which summarizes the most notable trends and advances in research on synthetic data generating models for ML solutions, with special attention to the preservation of privacy. On the one hand, a recent trend is observed towards the use of diffusion models and generative adversarial networks for the generation of synthetic data, while, on the other hand, it is concluded that there is no universal or objective way to evaluate fidelity and privacy. of the data generated synthetically by these generative models. Additionally, the importance of integrating generative models with other technologies that help preserve data privacy is highlighted, such as federated learning, differential privacy, and the use of secure data processing flows.

The next steps that we are starting to work on are the following:

• The definition and implementation of synthetic data generating models with privacy guarantees, based on diffusion models and generative adversarial networks.
• The compilation of metrics and methods to evaluate the data generated by these models with the objective of creating a universal fidelity and privacy evaluation methodology.
• The development and prototyping of a tool for the generation and evaluation of synthetic data that can be implemented in cloud-edge based architectures.

Collaboration with UC3m and support from the administration

A2-PRIVCOMP: Machine Learning with PRIVacy Restrictions for Computing at the Border, is a project funded within the UNICO R&D Cloud program within the framework of the Recovery, Transformation and Resilience Plan, financed by the European Union with NextGenerationEU funds. The coordinator of this project is Universidad Carlos III de Madrid, which launched a public tender to subcontract the development package of “Generation of synthetic data for machine learning solutions with privacy guarantees for cloud-edge infrastructure”, with Vicomtech being the successful bidder. of said tender.