Menu responsive
Home 9 Test Page

Challenges

9

Challenge #1 - Build efficient symmetric and asymmetric Functional Encryption

The first challenge that project Harpocrates considers is how to build efficient symmetric and asymmetric Functional Encryption (FE) schemes to support a wide range of statistical functions. Security of FE schemes must be improved by minimising the leakage associated with both the user’s query and the actual computation of the functions. To do so, we must ensure that our Functional Encryption schemes will be function- hiding in the sense that the CSP will output the correct result, without learning anything about the computed function. Another important challenge is designing a mechanism allowing users to explicitly specify the input of a function. In the standard Functional Encryption model, the function is applied to all of the users’ data. However, this may be extremely problematic in many cases when the function is not defined over some of the data.

9

Challenge #2 Inefficiencies of Differential Privacy (DP)

While Differential Privacy (DP) is a powerful tool for preserving the privacy of individuals, it currently suffers from important inefficiencies that will be addressed within the framework of HARPOCRATES. The first challenge that we will address is the design of a private encrypted database assuming a stronger threat model than the one presented in current literature. More precisely, in recent state-of-the-art approaches, the role of embedding well-calibrated noise to the ciphertexts, is given to the CSP. As a result, the security of such approaches is only satisfied under the assumption of an honest CSP. In HARPOCRATES, we will design schemes considering a malicious CSP. As a next step we will focus on the problem of minimising of the total accumulated noise after a sequence of updates in the database

9

Challenge #3 - Homomorphic Encryption (HE) inefficiencies

To bypass Homomorphic Encryption (HE) inefficiencies we aim at designing an Hybrid Homomorphic Encryption (HHE) scheme, by combining a symmetric-key encryption scheme with HE. However, symmetric schemes are not compatible with HE, mainly due to their large multiplicative depth. The first step of our research will revolve around comparing the compatibility of different symmetric schemes with HE. Consequently, we plan to design a symmetric scheme tailored around the needs of HE, with a strong focus on large integer arithmetic.

9

Challenge #4 - Design Machine Learning (ML) applications with enhanced security and stronger privacy

In HARPOCRATES, we aim to deploy Secure Multiparty Computation (MPC) protocols in combination with federated and split learning techniques, in an attempt to design Machine Learning (ML) applications with enhanced security and stronger privacy guarantees. Apart from that, in the field of Functional Encryption (FE), and in contrast with current state-of-the-art literature where an unrealistic fully trusted party generates and distributes functional decryption keys, MPC techniques can offer users the ability to generate those keys themselves and thus obviate the need of a fully trusted third party. Hence, in HARPOCRATES we will examine how to utilise MPC to eliminate the need for any trusted authority.

9

Challenge #5 Challenge with privacy-preserving Machine Learning

The main challenge with privacy-preserving Machine Learning (ML), is that Homomorphic Encryption (HE) and Functional Encryption (FE) schemes do not currently provide support for non-linear functions. To this end, we will focus on finding the best possible polynomial approximations for the activation functions used in ML. Apart from that, we seek to explore the possibility of designing privacy-preserving models for the classification of encrypted files (image, audio and video) – a problem we believe will make a real difference in providing guarantees to end users about their privacy.

9

Challenge #6 Challenges with Federated learning

Despite a promising outlook for analysing data in decentralised settings and federated data spaces, Federated Learning (FL) presents important challenges in three dimensions: data privacy, model confidentiality, and robustness to Byzantine attacks. In HARPOCRATES we will design an FL scheme combining existing approaches to protect data privacy (through multi-party secure aggregation), model confidentiality (with confidential computing) and Byzantine robustness.

Impact

    • Improved scalable and reliable privacy-preserving technologies for federated processing of personal data and their integration in real-world systems.

    • More user-friendly solutions for privacy-preserving processing of federated personal data registries by researchers. Improving privacy-preserving technologies for cyber threat intelligence and data sharing solution. Strengthened European ecosystem of open-source developers and researchers of privacy-preserving solutions.

    • Contribution to promotion of GDPR compliant European data spaces for digital services and research (in synergy with topic DATA-01-2021 of Horizon Europe Cluster 4)

    • Strengthened EU cybersecurity capacities and European Union sovereignty in digital technologies. More resilient digital infrastructures, systems and processes. Increased software, hardware and supply chain security. Secured disruptive technologies. Smart and quantifiable security assurance and certification shared across the EU.