Periodic Reporting for period 1 - HARPOCRATES (Federated Data Sharing and Analysis for Social Utility)
Periodo di rendicontazione: 2022-10-01 al 2024-03-31
Availability of large volumes of user data combined with tailored statistical analysis present a unique opportunity for organizations across the spectrum to adapt and finetune their services according to individual needs. Having shown remarkable results in analyzing user data, machine learning models attracted global adulation and are applied in a plethora of applications including medical diagnostics, pattern recognition, and threat intelligence. However, such service improvements and personalization based on user data analysis come at the heavy cost of privacy loss. Furthermore, practice showed that systems that use such models incorporate proxies that are often inexact, biased and often unfair. In HARPOCRATES, we focus on setting the foundations of digitally blind evaluation systems that will, by design, eliminate proxies such as geography, gender, race, and others and eventually have a tangible impact on building fairer, democratic and unbiased societies. To do so, we are designing several practical cryptographic schemes (Functional Encryption and Hybrid Homomorphic Encryption) for analyzing data in a privacy-preserving way. Besides processing statistical data in a privacy-preserving way, we also aim to enable a richer, more balanced and comprehensive approach where data analytics and cryptography go hand in hand with a shift towards increased privacy. In HARPOCRATES we first show how to effectively combine cryptography with the principles of differential privacy to secure and privatise databases. Next, we build privacy-preserving machine learning models able to classify encrypted data by performing high accuracy predictions directly on ciphertexts across federated data spaces. Finally, to demonstrate how these solutions respond to users’ needs, we implement two real-world cross-border data sharing scenarios related to health data analysis for sleep medicine and threat intelligence for local authorities.
To achive the above aims, the project target the following objectives:
1. Design efficient function-hiding Functionmal Encryption schemes to support privacy preserving data analytics.
2. Combine Functional Encryotion and Diffrerential Privacy for private encrypted databases.
3. Design a practical multi-client Hybrid Homomorphic Encryption scheme.
4. Build a Privacy Preserving Machime Learning framework by combining Functional Encryption and Hybrid Homomorphic Encryption.
5. Implement a Byzantine-robust Federated Learning scheme with data privacy guarantees.
6. Contribute to open science and reproducible research and develop two realistic cross-border demonstrator applications in the areas of sleep medicine and threat intelligence exchange for local authorities.
To achive the above aims, the project target the following objectives:
1. Design efficient function-hiding Functionmal Encryption schemes to support privacy preserving data analytics.
2. Combine Functional Encryotion and Diffrerential Privacy for private encrypted databases.
3. Design a practical multi-client Hybrid Homomorphic Encryption scheme.
4. Build a Privacy Preserving Machime Learning framework by combining Functional Encryption and Hybrid Homomorphic Encryption.
5. Implement a Byzantine-robust Federated Learning scheme with data privacy guarantees.
6. Contribute to open science and reproducible research and develop two realistic cross-border demonstrator applications in the areas of sleep medicine and threat intelligence exchange for local authorities.
During the first 18 months of the project, HARPOCRATES made significant advances towards all of its scientific objectives.
Towards Objective 1, the project defined a novel formal threat model that is applicable to any Functional Encryption (FE) scheme. It identified information that may leak during the run of FE schemes and made the first attempt to quantify the leakage in FE schemes. Additionally, HARPOCRATES has designed three Multi-Client FE schemes that cater to different technological potentials and use cases, including a lightweight multi-client FE scheme based on elliptic curves, a verifiable multi-client FE scheme tailored for inner products providing users and key curators with enhanced management of ciphertexts accessible by analysts, and a practical and novel FE scheme focused on the healthcare sector, particularly addressing the Sleep Medicine demonstrator. Towards Objective 2, the project combined Differential Privacy (DP) with FE mainly focusing on the utilisation of specific noise generation functions applicable in FE constructions to provide an even stronger privacy guarantee to users. Towards Objective 3, HARPOCRATES conducted a study of existing and widely used Homomorphic Encryption (HE) and Hybrid Homomorphic Encryption (HHE) schemas by analyzing several implementation aspects such as computational complexity, processing power, memory allocation, and security, as well as the ease of running a scheme. Additionally, it implemented a novel HHE library containing three HHE schemes (PASTA, HERA and Rubato). Having this library the project ran compatibility evaluation of the three HHE schemes. Towards Objective 4 HARPOCRATES developed a promising approach for private feature selection. This method utilizes low complexity bounds on mutual information to achieve efficient and secure feature selection using secure Multi-Party Computation (MPC). Additionally, it has designed several Privacy Preserving Machine Learning (PPML) protocols to support classification of encrypted data and started the design, implementation and evaluation of hybrid FE-HHE protocol for identifying objects in encrypted videos. Towards Objective 5 the project has evaluated existing secure aggregation schemes for Federated Learning (FL) parameter updates and started developing a novel MPC scheme to create a Privacy-Preserving Byzantine-robust Federated Learning service to offer global data privacy. Finally, towards Objective 6, the first versions of the HARPOCRATES demonstrators for analysing sleep medicine data and threat intelligence information between local authorities were implemented. The integration of these applications with the HARPOCRATES security services will commence in the second period of the project.
Towards Objective 1, the project defined a novel formal threat model that is applicable to any Functional Encryption (FE) scheme. It identified information that may leak during the run of FE schemes and made the first attempt to quantify the leakage in FE schemes. Additionally, HARPOCRATES has designed three Multi-Client FE schemes that cater to different technological potentials and use cases, including a lightweight multi-client FE scheme based on elliptic curves, a verifiable multi-client FE scheme tailored for inner products providing users and key curators with enhanced management of ciphertexts accessible by analysts, and a practical and novel FE scheme focused on the healthcare sector, particularly addressing the Sleep Medicine demonstrator. Towards Objective 2, the project combined Differential Privacy (DP) with FE mainly focusing on the utilisation of specific noise generation functions applicable in FE constructions to provide an even stronger privacy guarantee to users. Towards Objective 3, HARPOCRATES conducted a study of existing and widely used Homomorphic Encryption (HE) and Hybrid Homomorphic Encryption (HHE) schemas by analyzing several implementation aspects such as computational complexity, processing power, memory allocation, and security, as well as the ease of running a scheme. Additionally, it implemented a novel HHE library containing three HHE schemes (PASTA, HERA and Rubato). Having this library the project ran compatibility evaluation of the three HHE schemes. Towards Objective 4 HARPOCRATES developed a promising approach for private feature selection. This method utilizes low complexity bounds on mutual information to achieve efficient and secure feature selection using secure Multi-Party Computation (MPC). Additionally, it has designed several Privacy Preserving Machine Learning (PPML) protocols to support classification of encrypted data and started the design, implementation and evaluation of hybrid FE-HHE protocol for identifying objects in encrypted videos. Towards Objective 5 the project has evaluated existing secure aggregation schemes for Federated Learning (FL) parameter updates and started developing a novel MPC scheme to create a Privacy-Preserving Byzantine-robust Federated Learning service to offer global data privacy. Finally, towards Objective 6, the first versions of the HARPOCRATES demonstrators for analysing sleep medicine data and threat intelligence information between local authorities were implemented. The integration of these applications with the HARPOCRATES security services will commence in the second period of the project.
Although most of the technical work is still ongoing and its integration with the demonstrators is only scheduled for the second half of the project, HARPOCRATES has already made significant advancements beyond the state of the art. The project has designed the first FE scheme that is based on Elliptic Curve Cryptography (ECC), highlighting its novelty and potential impact. The efficiency of the scheme was demonstrated through theoretical and practical performance evaluation on two commercially available resource-constrained devices. The implemented verifiable multi-client FE scheme also goes beyond the state of the art. This achievement is significant as verification in FE has been overlooked in existing literature. Moreover, the project proposed a blockchain-based payment protocol between curators and analysts, facilitating the exchange of functional output for an appropriate amount. This payment protocol ensures fairness and atomicity of payments without relying on trust in the curator or other third parties. Furthermore, it utilises simple blockchain transactions, thereby reducing costs and improving the efficiency of the payment process.
The main exploitable result of the project will be a set of enablers (services), titled the HARPOCRATES Toolbox that currently includes six distinctive components (this number may change as the project evolves). The current HARPOCRATES enablers are the following: (1) Privacy Preserving Machine Learning through Hybrid Homomorphic Encryption, (2) Privacy Preserving Machine Learning through Functional Encryption, (3) Privacy Preserving Federated Learning, (4) Privacy Preserving Machine Learning with Differential Privacy-based data synthetization, (5) Robust Federated Learning with weight tuning, and (6) Privacy Preserving Machine Learning with Homomorphic Encryption. Based on these enablers, individual exploitation plans have already been outlined and will be further refined during the second half of the project.
The main exploitable result of the project will be a set of enablers (services), titled the HARPOCRATES Toolbox that currently includes six distinctive components (this number may change as the project evolves). The current HARPOCRATES enablers are the following: (1) Privacy Preserving Machine Learning through Hybrid Homomorphic Encryption, (2) Privacy Preserving Machine Learning through Functional Encryption, (3) Privacy Preserving Federated Learning, (4) Privacy Preserving Machine Learning with Differential Privacy-based data synthetization, (5) Robust Federated Learning with weight tuning, and (6) Privacy Preserving Machine Learning with Homomorphic Encryption. Based on these enablers, individual exploitation plans have already been outlined and will be further refined during the second half of the project.