Federated learning for digital twin applications: a privacy-preserving and low-latency approach

Jie Li, Dong WangPeerJ 2025 · DOI

TL;DR

This paper introduces FL-FedDT, a novel approach for federated learning (FL) in digital twin (DT) applications that enhances privacy and reduces latency. It achieves this by using an improved Paillier homomorphic encryption method and a trust-based global aggregation strategy to protect sensitive data while maintaining high training accuracy.

What Problem Does This Paper Address?

Digital twins (DTs) are increasingly used to create real-time models of physical systems for analysis and optimization. However, collecting and processing data from physical entities for DTs raises significant privacy concerns, as sensitive information can be leaked during transmission to servers, even from "honest but curious" providers. Federated learning (FL) offers a way to train models on decentralized data without directly sharing raw data, but the model updates themselves can still be vulnerable to malicious actors. Furthermore, variations in how local devices map data to the DT (mapping bias) and the presence of malicious clients can degrade the accuracy of FL models. This paper aims to tackle these challenges by providing a privacy-preserving and low-latency FL solution for DT applications.

Methods

The core of the proposed solution, FL-FedDT, lies in two main innovations:

  • Enhanced Paillier Homomorphic Encryption: The authors adapt the Paillier homomorphic encryption algorithm to reduce computational overhead. This involves introducing a new hyperparameter and pre-calculating intermediate values during key generation. This optimization significantly speeds up the encryption process, which is crucial for low-latency applications, without compromising security. Homomorphic encryption allows computations (like aggregation) to be performed directly on encrypted data, ensuring that the server never sees the raw model parameters.
  • Trusted Global Aggregation: To combat issues like mapping bias and malicious client updates, the paper implements a trust-based aggregation strategy. This method assigns varying weights to local device updates based on their learning quality and interaction records. Devices with higher reputation and more reliable updates receive greater weight in the global model aggregation, effectively mitigating the impact of unreliable or malicious nodes.

Datasets

The paper uses the MNIST dataset for training and evaluation.

Key Insights and Ideas

  • Homomorphic Encryption is Key for DT Privacy: The paper strongly advocates for homomorphic encryption as a superior privacy-preserving technique for FL in DT scenarios compared to methods like differential privacy, which can reduce accuracy, or artificial interference, which increases latency.
  • Optimizing Encryption for Speed: The core contribution of the enhanced Paillier encryption is demonstrating that significant latency reductions are possible by optimizing the encryption process, making FL more practical for real-time DT applications.
  • Trust is Crucial for Robust FL: The trust-based aggregation mechanism highlights that simply aggregating all updates is insufficient. Incorporating reputation and quality metrics is vital for building resilient FL systems, especially when dealing with potentially biased or malicious participants.
  • DT Bias is a Real Challenge: The paper acknowledges that DT mapping biases can impact FL accuracy and proposes a mechanism to account for these biases through weighted aggregation.

Key Results

  • Privacy and Accuracy Maintained: The FL-FedDT scheme achieves training accuracy comparable to baseline methods, demonstrating that enhanced privacy does not come at the cost of model performance.
  • Significant Latency Reduction: The enhanced Paillier encryption leads to a substantial reduction in FL iteration time compared to the original Paillier algorithm, particularly as key lengths increase. For example, the time difference for a 512-bit key is 20.387s, and for a 1,024-bit key, it's 119.387s.
  • Improved Communication Overhead: The proposed scheme shows advantages in communication overhead compared to traditional Paillier encryption schemes.
  • Robustness Against Malicious Clients: The trust-based aggregation effectively mitigates the impact of malicious updates, ensuring the global model's accuracy and reliability.
  • Randomness Equivalence: Randomness tests show that the enhanced Paillier algorithm generates keys with essentially identical randomness to the original, indicating equivalent security levels.

Limitations and Future Work

  • Resource Constraints: The authors acknowledge that resource-constrained edge devices (like temperature sensors) might struggle with the computational overhead of encryption tasks.
  • Larger-Scale Networks: Challenges remain in scaling the approach to larger networks and more complex heterogeneous environments.
  • Physical Layer Security: For highly constrained devices, exploring secure communication mechanisms at the physical layer is suggested.
  • Collaborative Interference Models: The paper proposes future work in exploring collaborative interference models to further optimize FL efficiency and secure data transmission.

Summary updated

Abstract

The digital twin (DT) concept has recently gained widespread application for mapping the state of physical entities, enabling real-time analysis, prediction, and optimization, thereby enhancing the management and control of physical systems. However, when sensitive information is extracted from physical entities, it faces potential leakage risks, as DT service providers are typically honest yet curious. Federated learning (FL) offers a new distributed learning paradigm that protects privacy by transmitting model updates from edge servers to local devices, allowing training on local datasets. Nevertheless, the training parameters communicated between local mobile devices and edge servers may contain raw data that malicious adversaries could exploit. Furthermore, variations in mapping bias across local devices and the presence of malicious clients can degrade FL training accuracy. To address these security and privacy threats, this paper proposes the FL-FedDT scheme-a privacy-preserving and low-latency FL method that employs an enhanced Paillier homomorphic encryption algorithm to safeguard the privacy of local device parameters without transmitting data to the server. Our approach introduces an improved Paillier encryption method with a new hyperparameter and pre-calculates multiple random intermediate values during the key generation stage, significantly reducing encryption time and thereby expediting model training. Additionally, we implement a trusted FL global aggregation method that incorporates learning quality and interaction records to identify and mitigate malicious updates, dynamically adjusting weights to counteract the threat of malicious clients. To evaluate the efficiency of our proposed scheme, we conducted extensive experiments, with results validating that our approach achieves training accuracy and security on par with baseline methods, while substantially reducing FL iteration time. This enhancement contributes to improved DT mapping and service quality for physical entities. (The code for this study is publicly available on GitHub at: https://github.com/fujianU/federated-learning. The URL address of the MNIST dataset is: https://gitcode.com/Resource-Bundle-Collection/d47b0/overview?utm_source=pan_gitcode&index=top&type=href&;.).
BibTeX
@inproceedings{li2025federated,
  author = {Jie Li and Dong Wang},
  title = {{Federated learning for digital twin applications: a privacy-preserving and low-latency approach}},
  booktitle = {PeerJ},
  year = {2025},
  doi = {10.7717/peerj-cs.2877},
}
Figure