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

Published as Paper 2026/2499 on the Cryptology ePrint Archive, the research arrives at a time when facial recognition has transitioned from science fiction into a ubiquitous utility. Today, facial biometrics are central to modern authentication systems, enterprise security frameworks, and highly personalized consumer services. From unlocking mobile devices and verifying digital banking transactions to streamlining airport security and securing physical corporate facilities, the technology is deeply embedded in daily life. However, this widespread adoption has been shadowed by severe and growing privacy risks.

At the heart of the controversy surrounding facial recognition is the inherent sensitivity of biometric data. Unlike a password or a cryptographic key, a person’s face cannot be easily changed if it is compromised. Traditional facial recognition pipelines typically require raw images or high-dimensional biometric templates—often referred to as embeddings—to be stored on central servers or transmitted across networks during the matching process. This centralized handling creates massive honeypots of sensitive information that are prime targets for malicious actors. Data breaches involving biometric templates expose individuals to unprecedented risks of identity theft, unauthorized surveillance, and long-term privacy violations. Furthermore, users are frequently forced to completely trust service providers with their most intimate physical characteristics, raising significant regulatory and ethical concerns under stringent privacy frameworks such as the European Union’s General Data Protection Regulation.

To confront these deep-seated vulnerabilities, the team from the University of Catania 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 the subsequent matching process entirely under secure multi-party computation, commonly known as MPC.

Secure multi-party computation is a subfield of cryptography that allows multiple parties to jointly compute a function over their private inputs while keeping those inputs completely hidden from all other participants. In the context of GhostFaceMPC, this cryptographic paradigm is leveraged to ensure that facial recognition can take place across all operational stages—encompassing feature extraction, secure database storage, and final identity matching—without ever exposing raw images or biometric embeddings at any point in the pipeline. Neither the server running the verification service nor any intermediate network observer gains access to plaintext visual data or mathematical representations of the user’s face.

Achieving practical privacy-preserving face recognition has historically been hampered by immense computational roadblocks. Deep convolutional neural networks, which form the backbone of modern accurate face recognition, rely on complex mathematical operations, non-linear activation functions, and floating-point arithmetic that are notoriously difficult and resource-intensive to evaluate efficiently within cryptographic circuits. Prior attempts to secure these systems often resulted in prohibitive performance degradation, making them entirely impractical for real-world deployment where speed and scalability are paramount.

To overcome these performance bottlenecks, Carnemolla, Catalano, and Di Raimondo introduced a novel, multi-party computation-friendly quantization of the GhostFaceNet model alongside targeted circuit optimizations. This strategic architectural refinement drastically reduces the computational overhead traditionally associated with cryptographic deep learning evaluation. By carefully tailoring the neural network architecture to align with the strengths of multi-party computation protocols, the researchers successfully bridged the gap between theoretical cryptographic security and real-world practicality, demonstrating that advanced biometric verification can be executed securely without demanding unrealistic computing resources.

The significance of the research extends far beyond academic theory, as the authors have committed to open-source transparency by making their implementation publicly available. The source code for GhostFaceMPC has been released to the research and developer community and is hosted on GitHub under the repository managed by the cryptography research group at the University of Catania, enabling independent auditing, verification, and further collaborative development.

As organizations and governments worldwide grapple with the dual imperatives of robust security and uncompromising privacy protection, developments like GhostFaceMPC represent a crucial technological stepping stone. By proving that advanced deep learning models can operate on encrypted data streams via multi-party computation without sacrificing accuracy or utility, the University of Catania team has outlined a promising path forward for the future of biometric authentication and identity management.