GhostFaceMPC: Privacy-Preserving Face Recognition with Multi-Party Computation

As facial recognition becomes an increasingly central pillar of modern authentication systems, institutional security frameworks, and personalized consumer services, its widespread deployment has continually collided with severe privacy concerns. Biometric identifiers such as facial geometry are fundamentally different from traditional passwords or cryptographic keys; they cannot be changed if compromised. Once raw facial images or sensitive biometric embeddings are exposed, leaked, or stolen, the implications for individual privacy are permanent and far-reaching. Centralized databases storing these vectors represent high-value targets for malicious actors, leaving organizations and individuals struggling to balance utility with uncompromising security.

To confront these formidable challenges head-on, the University of Catania research team developed GhostFaceMPC, heralded as the first end-to-end, privacy-preserving face recognition system capable of evaluating both a deep convolutional neural network feature extractor and subsequent template matching entirely under secure multi-party computation. Secure multi-party computation, commonly referred to as MPC, is a cryptographic subfield that allows multiple parties to jointly compute a function over their private inputs while keeping those inputs completely hidden from view. By leveraging this advanced cryptographic paradigm, GhostFaceMPC enables the secure processing of facial data across every single stage of the traditional face recognition pipeline.

The innovation spans from initial feature extraction and secure database storage to the final matching phase, ensuring that neither raw facial images nor sensitive biometric embeddings are exposed in the clear at any point during the process. In traditional architectures, even encrypted databases must eventually decrypt data to perform mathematical comparisons, creating a dangerous decryption bottleneck where sensitive vectors are vulnerable. GhostFaceMPC circumvents this vulnerability by performing computations directly on encrypted data shares distributed among multiple independent servers, ensuring that no single entity ever gains access to the complete biometric profile or the underlying imagery.

Achieving practical multi-party computation for deep learning models has historically been plagued by severe performance bottlenecks. Neural networks rely heavily on non-linear operations and complex mathematical transformations that are notoriously difficult and computationally expensive to evaluate within cryptographic circuits. To make the system viable for real-world application, the researchers introduced a novel, multi-party computation-friendly quantization of the GhostFaceNet model alongside specialized circuit optimizations. These tailored adjustments dramatically reduce the computational overhead and communication complexity traditionally associated with running deep convolutional neural networks over secret-shared data, bridging the gap between theoretical cryptography and practical systems engineering.

The significance of this development extends well beyond theoretical computer science, offering a tangible blueprint for industries bound by strict regulatory frameworks regarding consumer privacy and data protection, such as finance, healthcare, law enforcement, and enterprise access control. By decoupling the utility of biometric identification from the inherent risk of data centralization, GhostFaceMPC demonstrates that high-accuracy face recognition can coexist with absolute cryptographic privacy.

Demonstrating a commitment to open science, collaborative research, and verifiable security, the research team has made the complete source code and implementation details publicly available to the global cryptographic and security communities via GitHub under the repository managed by the cryptography research group at the University of Catania. This open-source release invites security researchers, developers, and auditors to inspect, test, and build upon the GhostFaceMPC framework, paving the way for wider adoption and iterative improvements in the rapidly evolving domain of privacy-preserving machine learning.