Addressing Membership Inference Attack in Federated Learning with Model Compression

Gergely Dániel Németh, Miguel Angel Lozano, Novi Quadrianto, Nuria OliverarXiv 2023 · arXiv · OpenReview

Abstract

Federated Learning (FL) has been proposed as a privacy-preserving solution for machine learning. However, recent works have reported that FL can leak private client data through membership inference attacks. In this paper, we show that the effectiveness of these attacks on the clients negatively correlates with the size of the client's datasets and model complexity. Based on this finding, we study the capabilities of model-agnostic Federated Learning to preserve privacy, as it enables the use of models of varying complexity in the clients. To systematically study this topic, we first propose a taxonomy of model-agnostic FL methods according to the strategies adopted by the clients to select the sub-models from the server's model. This taxonomy provides a framework for existing model-agnostic FL approaches and leads to the proposal of new FL methods to fill the gaps in the taxonomy. Next, we analyze the privacy-performance trade-off of all the model-agnostic FL architectures as per the proposed taxonomy when subjected to 3 different membership inference attacks on the CIFAR-10 and CIFAR-100 vision datasets. In our experiments, we find that randomness in the strategy used to select the server's sub-model to train the clients' models can control the clients' privacy while keeping competitive performance on the server's side.

Introduction

Federated Learning (FL) is frequently championed as a "privacy-by-design" framework because it allows multiple parties to collaboratively train a model without sharing their raw data. Instead, only model updates—gradients or parameters—are exchanged. However, this decentralization does not guarantee absolute privacy. Membership Inference Attacks (MIAs) have emerged as a significant threat, allowing adversaries to determine whether a specific data point was part of a client's local training set simply by analyzing the model's outputs or updates.

A growing area of interest is Heterogeneous Federated Learning, where clients possess varying hardware capabilities. In these scenarios, a central server might maintain a large, complex model, while individual clients (e.g., mobile devices) train smaller "sub-models" or compressed versions. While methods like HeteroFL or FDropout facilitate this "stitching" of updates, their impact on privacy remained largely unexplored. This research addresses this gap by investigating how model compression and sub-model selection strategies influence vulnerability to MIAs.

A Taxonomy of Model-Agnostic Federated Learning

To systematically evaluate the privacy-performance trade-off, the authors propose a novel taxonomy for model-agnostic FL. This framework categorizes how client-side sub-models are derived from the global server model across three critical dimensions:

  1. Client Grouping: This defines how clients are partitioned. They may all use the same subset of parameters (One group), be divided into clusters based on hardware (Several groups), or each utilize a unique subset (Unique).
  2. Dynamics: This refers to when the parameter subset is determined. Subsets can be Fixed at the start of training or Resampled (changed) in every communication round.
  3. Policy: This describes the selection logic. Parameters can be chosen as a contiguous block (Submatrix) or selected via a Random process.

By applying this taxonomy, the researchers identified seven new integration strategies—such as Global Fixed Matrix (GFM), Global Fixed Random (GFR), and One-group Submatrix Resampled (OSR)—to benchmark against established methods.

Experimental Setup and Attacks

The study evaluated these strategies using three benchmark datasets: CIFAR-10, CIFAR-100, and FEMNIST (Federated Extended MNIST). FEMNIST is particularly valuable as it represents real-world "non-IID" (non-identically and independently distributed) data, where each client's data reflects unique user characteristics.

To measure privacy leakage, the models were subjected to three distinct Membership Inference Attacks:

  • Yeom: A baseline attack utilizing prediction loss.
  • LiRA: A sophisticated attack that uses "shadow models" to simulate target behavior.
  • tMIA: A state-of-the-art attack that tracks loss changes over time via knowledge distillation.

Key Insights: The Power of Randomness

The central finding of this research is that randomness serves as a potent privacy booster. Strategies that introduce stochasticity into the sub-model selection process—specifically those that resample parameters every round (e.g., OSR or GSR)—significantly mitigate the effectiveness of MIAs.

The researchers propose a "Frequency Hypothesis": the more frequently a client trains on a specific set of parameters, the more the global model "memorizes" that client's unique data distribution. By frequently changing which parameters a client updates, the "fingerprint" of their private data is blurred across the global model, making membership inference much harder for an attacker.

Performance vs. Privacy Results

The experiments revealed several critical outcomes:

  • Privacy Gains: The proposed GFM and GFR methods maintained competitive accuracy with state-of-the-art baselines like HeteroFL while improving privacy protection (measured by AUC) by 0.5% to 1.0%.
  • Optimal Balancer: The OSM (One group, Submatrix, Resampled) method emerged as a top performer, offering the best balance between high server-side accuracy and robust client privacy.
  • The Impact of Scale: In federations dominated by "small" clients (those with limited hardware and smaller sub-models), the choice of integration strategy is paramount. The difference in privacy leakage between the best and worst strategies was up to three times higher in these scenarios compared to federations with "large" clients.

Conclusion and Future Directions

This work demonstrates that model compression in Heterogeneous FL is not just a necessity for hardware constraints but also a tunable knob for privacy. By strategically introducing randomness into how sub-models are selected and integrated, developers can significantly harden FL systems against membership inference.

However, challenges remain. Non-IID data distributions continue to pose a hurdle, sometimes causing server accuracy to drop as privacy increases. Future research will focus on refining these integration policies to handle extreme data skew and exploring the synergy between these architectural strategies and formal privacy frameworks like Differential Privacy.

Summary updated

Abstract

Federated Learning (FL) has been proposed as a privacy-preserving solution for machine learning. However, recent works have reported that FL can leak private client data through membership inference attacks. In this paper, we show that the effectiveness of these attacks on the clients negatively correlates with the size of the client's datasets and model complexity. Based on this finding, we study the capabilities of model-agnostic Federated Learning to preserve privacy, as it enables the use of models of varying complexity in the clients. To systematically study this topic, we first propose a taxonomy of model-agnostic FL methods according to the strategies adopted by the clients to select the sub-models from the server's model. This taxonomy provides a framework for existing model-agnostic FL approaches and leads to the proposal of new FL methods to fill the gaps in the taxonomy. Next, we analyze the privacy-performance trade-off of all the model-agnostic FL architectures as per the proposed taxonomy when subjected to 3 different membership inference attacks on the CIFAR-10 and CIFAR-100 vision datasets. In our experiments, we find that randomness in the strategy used to select the server's sub-model to train the clients' models can control the clients' privacy while keeping competitive performance on the server's side.
BibTeX
@misc{nemeth2023addressing,
  author = {Gergely Dániel Németh and Miguel Angel Lozano and Novi Quadrianto and Nuria Oliver},
  title = {{Addressing Membership Inference Attack in Federated Learning with Model Compression}},
  year = {2023},
  eprint = {2311.17750},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2311.17750},
}
Figure