Linkage on security, privacy and fairness in federated learning: New balances and new perspectives

Linlin Wang, Tianqing Zhu, Wanlei Zhou, Philip S. Yu

TL;DR

This survey paper explores the intricate interdependencies between security, privacy, and fairness in Federated Learning (FL). It identifies fairness as a critical "bridge" that mediates the trade-offs between privacy-preserving techniques and security defenses, offering a holistic perspective on building robust and equitable distributed systems.

What Problem Does This Paper Address?

Federated Learning (FL) was originally conceived to solve the "data island problem" by allowing multiple parties to train a shared model without exchanging raw data. However, the field has hit a wall: privacy, security, and fairness are often treated as independent goals.

The authors argue that this siloed approach is dangerous. For example, a technique used to protect privacy (like adding noise) might inadvertently make the model biased against minority groups. Conversely, a security defense designed to filter out malicious "outlier" updates might accidentally silence legitimate users with unique data. This paper tackles the "linkage" problem—how these three pillars interact and how to find a stable equilibrium between them.

Methods

As a comprehensive survey, the paper provides a new taxonomy and conceptual framework rather than a single new algorithm. Its core contribution is the "Fairness as a Bridge" model.

  • Taxonomy of Attacks: The authors categorize privacy attacks (e.g., gradient reconstruction, membership inference) and security threats (e.g., backdoor and poisoning attacks).
  • Linkage Analysis: They systematically map how defenses in one area impact the others. For instance, they analyze how Differential Privacy (DP) affects model fairness.
  • Fairness Categorization: They divide fairness-enhancing methods into three types:
  • Aggregation Algorithms: Optimizing how the central server combines client updates (e.g., FedProx, Ditto).
  • Incentive Mechanisms: Rewarding clients based on their contribution to encourage honesty and participation.
  • Personalization: Tailoring models to individual client needs to handle data heterogeneity.

Datasets

As this is a survey paper, the authors do not introduce a new dataset or conduct primary experiments. Instead, they review and synthesize results from hundreds of studies that utilize standard benchmarks in the FL community, including:

  • Image Classification: Used to demonstrate gradient leakage and label-flipping attacks.
  • IoT and Mobile Data: Used to discuss privacy-preserving applications in resource-constrained environments.
  • Healthcare and Banking Records: Cited as primary use cases where the privacy-security-fairness triad is most critical.

Key Insights and Ideas

  • The Gradient Sharing Paradox: The very mechanism that makes FL possible—sharing gradients—is the root cause of both privacy leaks (reconstructing data from gradients) and security vulnerabilities (poisoning the global model via malicious gradients).
  • The "Poor Get Poorer" Effect: One of the most striking insights is that Differential Privacy (DP) often disproportionately harms minority groups. Because DP adds noise to the training process, classes or users that are already difficult to classify (the "minority") see their accuracy drop much faster than the "majority" groups.
  • Defense Collateral Damage: Security defenses often rely on anomaly detection to find malicious actors. However, in a world of "Non-IID" (heterogeneous) data, a legitimate user with rare data looks exactly like an attacker. Filtering them out solves a security problem but creates a massive fairness problem.
  • Fairness as the Balancing Force: The authors posit that fairness can function as a conduit. By ensuring a model is fair, researchers can often identify where privacy or security measures have been over-applied or where they are failing.

Key Results

The paper summarizes the state of the field through several "Linkage" tables (Tables 2, 4, and 5):

  • Privacy vs. Fairness: Stricter privacy guarantees (lower epsilon in DP) almost universally lead to lower group fairness.
  • Security vs. Fairness: Backdoor attacks specifically target model fairness by making the model fail on specific "trigger" inputs while appearing normal otherwise.
  • Quantitative Trends: The authors note that bias in FL models typically grows at a much faster rate than the decline in overall accuracy when under attack or when privacy noise is added.

Limitations and Future Work

The authors acknowledge that while we understand these trade-offs exist, we lack a unified mathematical framework to optimize all three simultaneously.

Promising directions include: * Personalized FL: Developing methods that allow for local model customization to improve fairness without sacrificing global privacy. * Robust Incentive Design: Creating blockchain-based or reputation-based systems that can distinguish between a "malicious" client and a "unique" client. * Dynamic Equilibrium: Researching how to adjust privacy and security parameters in real-time based on the fairness metrics of the evolving global model.

Summary updated

Abstract

Federated learning is fast becoming a popular paradigm for applications involving mobile devices, banking systems, healthcare, and IoT systems. Hence, over the past five years, researchers have undertaken extensive studies on the privacy leaks, security threats, and fairness associated with these emerging models. For the most part, these three critical concepts have been studied in isolation; however, recent research has revealed that there may be an intricate interplay between them. For instance, some researchers have discovered that pursuing fairness may compromise privacy, or that efforts to enhance security can impact fairness. These emerging insights shed light on the fundamental connections between privacy, security, and fairness within federated learning, and, by delving deeper into these interconnections, we may be able to significantly augment research and development across the field. Consequently, the aim of this survey is to offer comprehensive descriptions of the privacy, security, and fairness issues in federated learning. Moreover, we analyze the complex relationships between these three dimensions of cyber safety and pinpoint the fundamental elements that influence each of them. We contend that there exists a trade-off between privacy and fairness and between security and gradient sharing. On this basis, fairness can function as a bridge between privacy and security to build models that are either more secure or more private. Building upon our observations, we identify the trade-offs between privacy and fairness and between security and fairness within the context of federated learning. The survey then concludes with promising directions for future research in this vanguard field.
BibTeX
@article{wang2025linkage,
  author = {Linlin Wang and Tianqing Zhu and Wanlei Zhou and Philip S. Yu},
  title = {{Linkage on security, privacy and fairness in federated learning: New balances and new perspectives}},
  journal = {Neural Networks},
  year = {2025},
  doi = {10.1016/j.neunet.2025.107874},
}
Figure