Dhiria: the technology

The technological engine of Dhiria stems from the ambition to constantly and significantly advance the scientific research. Achieving this goal requires addressing and solving the challenges the international scientific community poses. Thanks to proper investments, Dhiria, first in the market, achieved important results in terms of privacy-preserving computation, hence being able to bring this novel technology to the market.

More in detail, the research activity carried out in Dhiris allowed to design, develop and bring to the market machine and deep learning solutions able to process encrypted data in "as-a-service" manner. We emphasize that in all the phases of the processing data and results remain encrypted.  

This valuable result is the outcome of a scientific advance able to combine new neural network architectures and new training algorithms capable of preserving user privacy with innovative machine computation paradigms and deep learning provided in an "as-a-service" manner.

First in the market, DHIRIA has addressed and solved three important scientific and technological challenges:

• Algorithmic Challenge: Machine and deep learning models have been redesigned and suitably developed to operate on encrypted data.
• Technological Challenge: Privacy-preserving computing is characterized by high memory and computational demands. Hence, the machine and deep learning algorithms have been redesigned to reduce the computational load, memory occupation and energy consumption, while guaranteeing processing accuracy.
• Architectural challenge: in order to bring to market effective and efficient machine and deep learning solutions capable of operating on encrypted data, we implemented an "as-a-service" approach in which the processing is performed on high-performance systems in data centers or on Cloud.

The novel research solutions and the main scientific results achieved by Dhiria have been published in the main international journals and conferences:

• A. Falcetta, M. Roveri, “Privacy-preserving machine learning with homomorphic encryption: an introduction”, IEEE Computational Intelligence Magazine, August, 2022 [Q1 SCIMAGO – Computer Science]. 
• M. Gambella, A. Falcetta, M. Roveri, “CNAS: Constrained Neural Architecture Search”, Accepted at the 2022 IEEE International Conference on Systems, Man and Cybernetics (IEEE SMC 2022),Prague, 2022.
• A. Falcetta, M. Roveri, “TIMEX: an Automatic Framework for Time-Series Forecasting-as-a-Service”, Accepted at the 6th International Workshop on Automation in Machine Learning, held in conjunction with the SIGKDD Conference on Knowledge Discovery and Data Mining conference (KDD2022),Washington DC, 2022.
• A. Falcetta, M. Roveri, “Privacy-preserving time series prediction with temporal convolutional neural networks”, in Proc. 2022 International Joint Conference on Neural Network (IJCNN2022),Padova, Italy, 2022.
• S. Disabato, A. Falcetta, A. Mongelluzzo, M. Roveri, “A privacy-preserving distributed architecture for deep-learning-as-a-service”, in Proc. 2020 International Joint Conference on Neural Networks (IJCNN),Glasgow, 2020.

These publications show the novelty and the effectiveness of the research carried out in Dhiria in the field of privacy-preserving machine and deep learning "as-a-service"

A further relevant point deserves our attention: the scientific results in privacy-preserving computation achieved by Dhiria represent a valuable asset also for the processing of plain (un-encrypted) data.

Indeed, processing encrypted data requires dealing with memory- and computational-demanding computation and this pushes us constantly towards optimizing every component of our platform.

The novel solutions achieved in the algorithmic, technological and architectural fields developed by Dhiria provide a relevant competitive advantage, compared to other solutions present in the market, in bringing to the customers machine and deep learning as-a-service solutions that operate on plain data, guaranteeing better performance and lower costs of processing.