Privacy Preserving Multi-Party-Computing (MPC):
Share data for multi-party Machine Learnig applications and analysis without disclosing access to sensitive data
Privacy Preserving Machine Learning
The joint analysis and use of data from different parties, e.g. from machines, usually fails due to issues of confidentiality, competition and direct competition. The “Trusted Data Hub”, a novel combination of hardware and software technologies, provides the solution to protect company secrets while sharing data for AI-based analysis.
Data remains private, public analyses become possible
“Trusted Hub Solution” offers manufacturer-independent, data sovereign machine learning (PPML) and multi-party computing (MPC). PPML enables the training of the artificial intelligence (AI) model and machine learning (ML) in a way that preserves the privacy of the data.
Analyse data without disclosing it
The MPC allows institutions to carry out analyses of private data held by several other institutions without ever disclosing the data. In other words, MPC is a technique where calculations and data analysis are performed with input from various parties. The parties are not given insight into the data provided by the other parties – only the results of the data analysis are shared.
Guarantee of confidentiality
The Trusted Data Hub solution guarantees that the participating parties only receive the results, but no information about the input of others.
Use cases in vertical domains
Many applications in vertical domains can benefit from the Trusted Data Hub. An important example is the area of predictive maintenance. To implement this, machine manufacturers usually need access to raw data from multiple users. At the same time, data owners do not simply want to disclose their trade secrets or machine-generated data (MGD).
Data control and data sovereignty
The Trusted Data Hub enables analysis of the combined vulnerable data from different companies without losing control of their data. At the same time, machine manufacturers can significantly improve their products and services by evaluating or analyzing the combined data of their customers. For example, machine owners benefit from improved machine availability and machine settings or from reduced wear and tear.
In contrast to all other existing solutions, the Trusted Hub is completely independent of the analytic tools used. While offering the same accuracy and runtime, the Trusted Hub supports not only Python, but also all other ML frameworks, such as Rapidminer, Knime, H2O and Matlab.
Easy integration with full ML compatibility
The Trusted Hub is relatively easy to integrate into existing systems, so it addresses the issue of data protection without impacting the application and customer experience. The Trusted Hub supports all data types (i.e. numeric and categorical data) and all ML functions (i.e. linear and non-linear functions) and is therefore compatible with all types of ML algorithms.
Uncompromising data sovereignty
Furthermore, the Trusted Hub not only protects the data, but can also protect the ML algorithm itself, i.e. only the creator of the ML algorithm can access the code.
The implementation and setup of the Trusted Hub is very fast and uncomplicated. Furthermore, no individual adaptation is necessary.
|Privacy Preserving and Multi Party Computation Technique||Trusted Data Hub |
|Additive key/secret Sharing||Homomorphic Encryption (HE)||Differential privacy (DP)||Federated Learning (FL)|
|Centralized/Decentralized||Centralized||Decentralized with a Trusted Server||Centralized||Centralized||Decentralized with a Trusted Server|
|Multi Party Commputation Support||Yes||Yes||No||Yes||Yes|
|Data Privacy Level||High||High||Very High||High||High|
|ML Algorithm Privacy Level||Yes||No||Yes||Yes||No|
|Runtime & Scalability||High||Medium||Very low||Medium||Medium|
|Accuracy||High||Medium||Medium||Medium||Medium to High|
|Programming Language & ML Framework||All Frameworks||Python||Vendor-dependent||Based on Python and Tensorflow||Tensorflow and PySyft|
|Supported Data Types||Numerical and categorical||Numerical data||Numerical data||Numerical and categorical||Numerical and categorical|
|Supported Functions||All functions||Not all||Not all||All Functions||All functions|
© ADVANEO GmbH
Alle hier gezeigten Inhalte und Grafiken unterliegen dem Urheberrecht und dürfen ohne unsere Erlaubnis nicht kopiert oder verwendet werden.