Fluent: Round-efficient Secure Aggregation for Private Federated Learning

Xincheng Li, Jianting Ning, Geong Sen Poh, Leo Yu Zhang, Xinchun Yin, Tianwei ZhangComput. Stand. Interfaces 2024 · DOI · OpenReview

TL;DR

This paper introduces "Fluent," a novel secure aggregation scheme for federated learning that significantly reduces communication rounds and overhead. By optimizing secret sharing and combining consistency checks with unmasking, Fluent achieves greater efficiency and scalability, making private federated learning more practical for real-world applications.

What Problem Does This Paper Address?

Federated learning (FL) allows training machine learning models on decentralized data while preserving user privacy. However, existing FL methods are vulnerable to privacy attacks like inference and inversion. Secure aggregation schemes aim to mitigate these threats by encrypting and combining model updates. Current single-server secure aggregation methods, while promising, suffer from high communication round complexity and computational overhead, especially for geographically distributed clients. This latency and cost hinder the practical deployment of FL. Existing solutions often require multiple communication rounds for tasks like consistency checks and mask reconstruction, leading to significant delays.

Methods

Fluent introduces two key innovations to address the limitations of existing secure aggregation schemes:

  • One-time Handshake and Reusable Secret Sharing: Unlike previous methods that require frequent handshakes and secret sharing for masks in each training iteration, Fluent performs these operations only once during a pre-round phase. Decryptors can then reuse these secret shares across multiple aggregation iterations. This significantly reduces computational and communication costs during the main training loop. The scheme leverages the linearity of Shamir's secret sharing and a one-way function to ensure privacy.
  • One-Round Consistency Check and Unmasking: Fluent combines the consistency check (to detect malicious servers manipulating dropout information) and the gradient unmasking process into a single logical step. This eliminates an entire communication round compared to prior work. Decryptors provide encrypted shares, and the server can decrypt them to reconstruct seeds and unmask gradients only after the consistency check is successful. This optimization is crucial for reducing latency in distributed settings.

Fluent also introduces "Fluent-Dynamic" with a new participant selection algorithm and a multilevel threshold secret sharing scheme to support dynamic client and committee joining, enhancing system flexibility and scalability.

Datasets

The paper does not explicitly mention specific datasets used for training the models in the experimental evaluation. However, it states that they train a convolutional neural network (CNN) on the MNIST and CIFAR-10 datasets from Torchvision for practical use cases.

Key Insights and Ideas

  • Reusability is Key: The core insight is that secret shares for masks don't need to be regenerated every iteration. Reusing these shares from an initial setup dramatically cuts down on communication and computation.
  • Combining Steps for Efficiency: Merging the consistency check and unmasking into a single round is a clever way to shave off valuable communication time, which is a major bottleneck in FL.
  • Dynamic Participation is Crucial for Scalability: The ability for clients and decryptors to join and leave dynamically makes the system much more adaptable to real-world scenarios where participants might not be static.
  • Trade-offs are Managed: While Fluent reduces communication rounds, it slightly increases computational cost for the server and decryptors. However, this trade-off is deemed acceptable due to the significant gains in communication efficiency, especially for resource-constrained clients.

Key Results

Experimental results show that Fluent significantly outperforms existing state-of-the-art schemes:

  • Client Efficiency: Fluent improves computational cost by at least 75% and communication overhead by at least 25% for normal clients.
  • Server Efficiency: Fluent reduces communication overhead for the server.
  • Round Efficiency: Fluent achieves the fewest communication rounds (two in the collection phase) in the malicious server setting, compared to at least three rounds in existing schemes.
  • Dropout Robustness: Fluent exhibits superior performance to Flamingo in terms of dropout robustness.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that Fluent requires a longer waiting time for decryptors to compute encrypted shares, potentially leading to slower overall model training compared to schemes with less complex decryption.
  • While Fluent offloads computation to the server and decryptors, this can still be a burden for very resource-limited servers or decryptor nodes.
  • Future Work:
    • The authors plan to address more security threats in FL, specifically focusing on defenses against poisoning attacks (both untargeted and targeted backdoor attacks) to ensure the integrity and correctness of the global model. This is noted as a challenging area due to the stealthy nature of such attacks.

Summary updated

Abstract

Federated learning (FL) facilitates collaborative training of machine learning models among a large number of clients while safeguarding the privacy of their local datasets. However, FL remains susceptible to vulnerabilities such as privacy inference and inversion attacks. Single-server secure aggregation schemes were proposed to address these threats. Nonetheless, they encounter practical constraints due to their round and communication complexities. This work introduces Fluent, a round and communication-efficient secure aggregation scheme for private FL. Fluent has several improvements compared to state-of-the-art solutions like Bell et al. (CCS 2020) and Ma et al. (SP 2023): (1) it eliminates frequent handshakes and secret sharing operations by efficiently reusing the shares across multiple training iterations without leaking any private information; (2) it accomplishes both the consistency check and gradient unmasking in one logical step, thereby reducing another round of communication. With these innovations, Fluent achieves the fewest communication rounds (i.e., two in the collection phase) in the malicious server setting, in contrast to at least three rounds in existing schemes. This significantly minimizes the latency for geographically distributed clients; (3) Fluent also introduces Fluent-Dynamic with a participant selection algorithm and an alternative secret sharing scheme. This can facilitate dynamic client joining and enhance the system flexibility and scalability. We implemented Fluent and compared it with existing solutions. Experimental results show that Fluent improves the computational cost by at least 75% and communication overhead by at least 25% for normal clients. Fluent also reduces the communication overhead for the server at the expense of a marginal increase in computational cost.
BibTeX
@article{li2024fluent,
  author = {Xincheng Li and Jianting Ning and Geong Sen Poh and Leo Yu Zhang and Xinchun Yin and Tianwei Zhang},
  title = {{Fluent: Round-efficient Secure Aggregation for Private Federated Learning}},
  journal = {Comput. Stand. Interfaces},
  year = {2024},
  doi = {10.1016/j.csi.2025.104096},
}
Figure