MQFL-FHE: Multimodal Quantum Federated Learning Framework with Fully Homomorphic Encryption α \alpha These authors contributed equally to this work.
Siddhant Dutta 1 α \alpha , Nouhaila Innan 2 3 α \alpha , Sadok Ben Yahia 4 5 , Muhammad Shafique 2 3 , and David E. Bernal Neira 6 · 2024
Read MQFL-FHE: Multimodal Quantum Federated Learning Framework with Fully Homomorphic Encryption α \alpha These authors contributed equally to this work. by Siddhant Dutta 1 α \alpha , Nouhaila Innan 2 3 α \alpha , Sadok Ben Yahia 4 5 , Muhammad Shafique 2 3 , and David E. Bernal Neira 6 (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-01858.txt.
Introduction
Siddhant Dutta and co-authors present a 2024 framework that joins quantum computing with privacy-locked group learning. In federated learning, many sites train one model without sharing raw data, and full homomorphic encryption lets them merge updates while still coded. The catch: coded merging often dulls the merged model. Their design adds a quantum mixture of experts that reads several data types at once.
On mixed data sets, and on a hard case pairing gene data with brain scans, the quantum experts lifted accuracy, above all for rare classes. The quantum step also offset the accuracy lost to encryption. The authors offer the frame as a way to keep both privacy and performance.
Catalog record
| CREATOR | Siddhant Dutta 1 α \alpha , Nouhaila Innan 2 3 α \alpha , Sadok Ben Yahia 4 5 , Muhammad Shafique 2 3 , and David E. Bernal Neira 6 |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-01858
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
MQFL-FHE: Multimodal Quantum Federated Learning Framework with Fully Homomorphic Encryption α \alpha These authors contributed equally to this work. Siddhant Dutta 1 α \alpha , Nouhaila Innan 2 3 α \alpha , Sadok Ben Yahia 4 5 , Muhammad Shafique 2 3 , and David E. Bernal Neira 6 Abstract The integration of fully homomorphic encryption (FHE) in federated learning (FL) has led to significant advances in data privacy. However, during the aggregation phase, it often results in performance degradation of the aggregated model, hindering the development of robust representational generalization. In this work, we propose a novel multimodal quantum federated learning framework that utilizes quantum computing to counteract the performance drop resulting from FHE. For the first time in FL, our framework combines a multimodal quantum mixture of experts (MQMoE) model with FHE, incorporating multimodal datasets for enriched representation and task-specific learning. Our MQMoE framework enhances performance on multimodal datasets and combined genomics and brain MRI scans, especially for underrepresented categories. Our results also demonstrate that the quantum-enhanced approach mitigates the performance degradation associated with FHE and improves classification accuracy across diverse datasets, validating the potential of quantum interventions in enhancing privacy in FL. I Introduction Federated Learning (FL) has emerged as a powerful paradigm for collaboratively training machine learning models across decentralized data sources ( ( McMahan et al. 2023 ) ), allowing multiple clients to enhance model performance while preserving data privacy and complying with regulations such as the general data protection regulation (GDPR). FL alone is insufficient to fully protect data privacy, as the global model and updates shared between clients and the central server remain vulnerable to inference attacks. For instance, membership inference attacks use the model parameters or gradients to deduce whether a specific data sample was part of the training dataset ( ( Shokri et al. 2017 ) ). Similarly, attribute inference attacks attempt to uncover sensitive attributes of individual data samples by exploiting correlations between model updates and data features ( ( Melis et al. 2019 ) ). Fully homomorphic encryption (FHE) was proposed by ( Gentry 2009 ) to mitigate vulnerabilities by allowing computations on encrypted data without decryption, ensuring that sensitive information remains secure even during model training and aggregation within new gen FL. It works well against gradient leakage attacks , in which hackers try to recover private information by looking at the gradients that are sent back and forth during training ( ( Phong et al. 2018 ) ). Adding FHE to FL adds a lot of extra work to computers, causing delays and less accurate models during the aggregation phase ( ( Zhang et al. 2023 ) ). Consequently, while FHE protects data privacy, it can hinder the overall performance of FL systems, which is crucial for applications requiring high accuracy or real-time predictions. Despite the growing interest in multimodal FL, which aims to leverage various types of data from multiple clients, there exists a gap in research that addresses the unique challenges posed by FHE in this context ( ( Gong et al. 2024a ) ). Furthermore, the application of quantum computing (QC) to enhance FL frameworks has been largely unexplored. QC has the potential to accelerate specific computations, which could alleviate performance bottlenecks introduced by FHE ( ( Dutta et al. 2024 ) ). To mitigate the performance degradation that occurs during model aggregation, we propose a novel integration of QC with FHE in FL, taking into account the nature of the problem. Our method employs QC to counteract the model degradation resulting from FHE, allowing for an improved collection of model updates while protecting privacy, and for the first time, we present a novel multimodal quantum mixture of experts (MQMoE) model within the FL framework that employs FHE. We design this architecture to handle diverse multimodal data with MoE ( ( Yu et al. 2023a ) ) from various clients, enhancing representational and task-specific learning performance. The main contributions, as illustrated in Fig. 1 , are as follows: •Conducting a comprehensive range of single-modality experiments, from classical centralized approaches to FHE-based quantum federated learning (QFL), demonstrating how QC can mitigate model accuracy degradation caused by CKKSTensor-based FHE. •Developing a novel algorithmic framework for seamlessly integrating multimodal datasets with QC and FL while maintaining FHE constraints. •Proposing a novel MQMoE in the FL-FHE setup that leverages quantum layer outputs in its gating mechanism, achieving enhanced representational generalizability for task-specific learning. •Validating our approach through experiments in the biological domain, where we create a biological MQMoE incorporating two distinct expert types specifically designed for handling sensitive medical information (genomics and brain magnetic resonance imaging (MRI) scans). Fig. 1: Overview of our novel contributions. II Related Works QC in FL: ( Innan et al. 2024a ) introduced the federated quantum neural network (FedQNN), which integrates quantum machine learning (QML) with QFL principles ( ( Biamonte et al. 2017 ) ). The experiments were conducted across genomics and healthcare datasets, laying the groundwork for investigations into quantum-enhanced learning methods in privacy-sensitive domains. This has been further advanced by efforts in QFL, which are creating new avenues in QML. In particular, studies like ( Chen and Yoo 2021 ; Chehimi and Saad 2022 ) have emphasized using QFL in quantum systems and decentralized data frameworks. Applications in healthcare and other domains, using both quantum hardware and classical simulations, have shown promising results ( ( Lusnig et al. 2024 ; Bhatia and Bernal Neira 2024 ; Innan et al. 2024b ; Javeed et al. 2024 ; Bhatia et al. 2023 ; Kais et al. 2023 ) ). Despite this progress, challenges related to resource allocation and efficient implementation persist, as noted by ( Larasati et al. 2022 ; Ren et al. 2023 ; Gurung et al. 2023 ) , highlighting ongoing efforts to address these issues. Privacy preservation in FL: ( Rahulamathavan et al. 2023 ) proposed FheFL, which incorporates a multi-key additive homomorphic encryption scheme to mitigate privacy risks associated with gradient sharing in FL. The experimental results confirmed that FheFL maintains model accuracy while providing robust security against potential data breaches, emphasizing the necessity of integrating FHE techniques into FL frameworks to enhance data privacy. FedSHE was created by ( Pan et al. 2024 ) . It is a useful FL scheme that uses adaptive segmented CKKS homomorphic encryption to keep gradients safe from leakage attacks. Their analysis of security parameters underscores the trade-offs between correctness, efficiency, and security in FL environments. Multimodal data handling in FL: ( Yu et al. 2023b ) developed CreamFL, a framework designed for multimodal FL. This method enables clients with heterogeneous model architectures to contribute to a unified global model without sharing their private data. By implementing a contrastive representation ensemble strategy, CreamFL effectively addresses the challenges posed by modality discrepancies and task diversities. ( Gong et al. 2024b ) further advanced this area by introducing a multimodal vertical FL framework utilizing bivariate Taylor series expansion for encryption, thus eliminating the reliance on third-party encryption providers. Unlike the aspects discussed in previous literature, our methodology tackles the problem of managing multimodal data while reducing performance degradation due to FHE. We propose a seamless algorithm that integrates QC techniques with multimodal FL setups