FedRand: Enhancing Privacy in Federated Learning with Randomized LoRA Subparameter Updates

Sangwoo Park, Seanie Lee, Byungjoo Kim, Sung Ju Hwang

TL;DR

This paper introduces FedRand, a novel federated learning approach that enhances privacy for vision-language models (VLMs) by selectively sharing only a portion of the model's parameters during training. FedRand significantly improves resistance to membership inference attacks, a common privacy vulnerability in federated learning, while maintaining comparable performance to existing methods.

What Problem Does This Paper Address?

Federated Learning (FL) is a powerful technique for training machine learning models across decentralized data sources without directly accessing sensitive user data. However, even with FL, privacy remains a concern. When models, especially large ones like vision-language models (VLMs), are trained and their parameters are sent to a central server for aggregation, they can inadvertently memorize parts of the training data. This memorization makes them vulnerable to "membership inference attacks" (MIAs), where an attacker tries to determine if a specific data point was part of the training set. This is particularly problematic for VLMs due to their capacity to memorize complex information, posing risks in sensitive domains like healthcare or finance. Existing FL methods, while protecting raw data, still expose enough of the model's learned parameters to potentially leak private information.

Methods

FedRand tackles this privacy challenge by leveraging Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning technique. Instead of sending all updated LoRA weights from a client to the server, FedRand introduces a randomization strategy:

  1. Selective Parameter Sharing: At each training round, a client receives the full set of LoRA weights (matrices A and B) from the server. The client then randomly chooses to update either matrix A or matrix B, while keeping the other matrix as a "client-specific private parameter," initialized with its previous round's values.
  2. Partial Upload: Only the non-private (updated) LoRA parameters are sent back to the central server for aggregation. The private parameters remain on the client.
  3. Aggregation: The server aggregates the received non-private parameters to update its global LoRA weights.

This approach ensures that the central server never sees the complete set of a client's updated LoRA parameters, making it much harder for an attacker to reconstruct the client's specific model and infer membership.

Datasets

The paper uses standard datasets for evaluation:

  • ScienceQA: A multiple-choice visual question answering dataset focused on science curricula.
  • MSCOCO: A large-scale dataset for image recognition, segmentation, and captioning.
  • NoCaps: A dataset designed to evaluate out-of-distribution (OOD) generalization for image captioning, used here to test robustness against membership inference attacks.

The datasets were split using a Dirichlet distribution to simulate non-IID (non-independent and identically distributed) scenarios, common in real-world federated learning.

Key Insights and Ideas

  • LoRA as a Privacy Enabler: The paper demonstrates that LoRA, primarily known for efficiency, can be a powerful tool for enhancing privacy in federated learning by allowing for granular control over parameter sharing.
  • Randomization is Key: The core innovation is the random selection of which LoRA subparameter (A or B) to keep private. This randomness prevents attackers from consistently targeting specific parts of the model, making reconstruction much harder.
  • Trade-off Management: FedRand successfully navigates the typical trade-off between privacy and performance. It significantly boosts privacy against MIAs without a substantial drop in model accuracy, unlike some other privacy-focused methods.
  • Communication Efficiency: By sending only a subset of parameters, FedRand also offers a reduction in communication costs, a crucial aspect of federated learning.

Key Results

  • Privacy Enhancement: FedRand significantly improves robustness against membership inference attacks (MIA), as measured by AUROC scores, outperforming baselines like FedAvg, FedPer, and FedPara. For instance, in Table 4, FedRand (server) achieves a lower MaxRényi-10% score (indicating better robustness) than FedAvg (server).
  • Performance Comparability: FedRand achieves accuracy comparable to FedAvg (the oracle method that communicates full LoRA weights) on tasks like VQA and image captioning. For example, on the ScienceQA dataset, FedRand (server) achieves 80.12% accuracy, very close to FedAvg's 81.50%.
  • Communication Cost Reduction: Figure 3 shows that FedRand reduces the communication cost by approximately 25% per round compared to FedAvg, while maintaining similar accuracy.

Limitations and Future Work

  • Not a Complete Solution: The authors acknowledge that FedRand doesn't eliminate all privacy risks and that more sophisticated attacks might still be possible.
  • Hyperparameter Sensitivity: While the paper explores some hyperparameter choices (like the probability p of selecting A), optimal settings might require further tuning.
  • Future Directions: The authors suggest exploring random sub-layer selection for training or quantizing client parameters to further enhance security. They also note that while FedRand addresses privacy, it doesn't directly tackle fairness or bias in FL.

Summary updated

Abstract

Federated Learning (FL) is a widely used framework for training models in a decentralized manner, ensuring that the central server does not have direct access to data from local clients. However, this approach may still fail to fully preserve data privacy, as models from local clients are exposed to the central server during the aggregation process. This issue becomes even more critical when training vision-language models (VLMs) with FL, as VLMs can easily memorize training data instances, making them vulnerable to membership inference attacks (MIAs). To address this challenge, we propose the FedRand framework, which avoids disclosing the full set of client parameters. In this framework, each client randomly selects subparameters of Low-Rank Adaptation (LoRA) from the server and keeps the remaining counterparts of the LoRA weights as private parameters. After training both parameters on the client's private dataset, only the non-private client parameters are sent back to the server for aggregation. This approach mitigates the risk of exposing client-side VLM parameters, thereby enhancing data privacy. We empirically validate that FedRand improves robustness against MIAs compared to relevant baselines while achieving accuracy comparable to methods that communicate full LoRA parameters across several benchmark datasets.
BibTeX
@misc{park2025fedrand,
  author = {Sangwoo Park and Seanie Lee and Byungjoo Kim and Sung Ju Hwang},
  title = {{FedRand: Enhancing Privacy in Federated Learning with Randomized LoRA Subparameter Updates}},
  year = {2025},
  eprint = {2503.07216},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.07216},
}
Figure