SHFL: Secure Hierarchical Federated Learning Framework for Edge Networks

Omid Tavallaie, Kanchana Thilakarathna, Suranga Seneviratne, Aruna Seneviratne, Albert Y. ZomayaarXiv 2024 · DOI · arXiv

Abstract

Federated Learning (FL) enables decentralized AI training while preserving data privacy, but it remains vulnerable to poisoning attacks. As networks evolve into hierarchical structures with edge servers, traditional flat-network defenses become inadequate. This paper introduces SHFL, a secure hierarchical framework employing a two-tier defense: edge-level client selection and cloud-level convex optimization. SHFL effectively mitigates model and data poisoning attacks, maintaining high accuracy even in non-IID data environments where traditional methods fail.

Introduction: The Vulnerability of the Edge

Standard FL architectures are susceptible to malicious clients who perform model poisoning (sending corrupted gradients) or data poisoning (label-flipping). While defenses exist for simple client-server models, modern IoT deployments increasingly rely on Hierarchical Federated Learning (HFL). In HFL, edge servers act as intermediaries, aggregating local updates before communicating with the cloud.

Existing defenses often struggle to distinguish between malicious activity and the natural variance of non-IID (heterogeneous) data. This often leads to the accidental exclusion of legitimate users whose data is simply unique. The challenge lies in filtering out attackers without silencing diverse, benign voices in the network.

The SHFL Framework

SHFL addresses these challenges through a novel two-stage defense mechanism integrated into the hierarchical pipeline.

The SHFL architecture illustrates the multi-tier process: (1) local model upload, (2) edge aggregation, (3) edge model upload, (4) cloud aggregation, (5) global model download to edge, and (6) global model download to IoT devices.

1. Edge-Level Client Selection

Edge servers compute a trust metric for each IoT device based on the $L_2$-norm (Euclidean distance) between the local update and the current global model. By filtering out extreme outliers and then randomly selecting from the remaining "trusted" pool, SHFL prevents overfitting to a narrow user base while neutralizing obvious attackers. This randomness is a key innovation for maintaining fairness and model generalization.

2. Cloud-Level Aggregation (AGR)

At the cloud tier, SHFL replaces simple averaging with a convex optimization approach. Using Karush-Kuhn-Tucker (KKT) conditions, the cloud dynamically assigns weights to edge models based on data volume and proximity to the global state. This allows the system to mathematically "mute" suspicious edge contributions while prioritizing reliable, data-rich ones.

Experimental Evaluation

The framework was tested against standard benchmarks (MNIST, Fashion-MNIST, CIFAR-10) in both IID and non-IID scenarios. The primary threats modeled were Projected Gradient Ascent (PGA)—a sophisticated model poisoning attack—and Label Flipping (LF).

Training accuracy fluctuations under varying intensities of Projected Gradient Ascent (PGA) attacks, highlighting the instability that SHFL aims to mitigate.

Key Results

  • Resilience to Poisoning: In non-IID MNIST tests with 10 attackers, SHFL maintained over 70% accuracy, significantly outperforming FedAvg and Trimmed-mean, which plummeted to below 30%.
  • Stability: SHFL demonstrated superior stability in IID settings, maintaining over 94% accuracy against 10 PGA attackers, whereas competing methods exhibited high volatility across training rounds.
  • Low Performance Overhead: SHFL achieves accuracy comparable to standard FedAvg in non-adversarial environments, ensuring there is no "security tax" on performance during normal operation.

Conclusion and Future Directions

SHFL provides a robust defense for hierarchical edge networks by combining distance-based filtering with mathematical optimization. Future work will focus on refining attacker estimation and extending the framework to defend against targeted backdoor attacks, where specific inputs are misclassified without degrading overall model accuracy. Additionally, the impact of these calculations on the battery life of resource-constrained IoT devices remains an area for further empirical study.

Summary updated

BibTeX
@misc{tavallaie2024shfl,
  author = {Omid Tavallaie and Kanchana Thilakarathna and Suranga Seneviratne and Aruna Seneviratne and Albert Y. Zomaya},
  title = {{SHFL: Secure Hierarchical Federated Learning Framework for Edge Networks}},
  year = {2024},
  eprint = {2409.15067},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.15067},
}
Figure