A Hybrid Federated–Incremental Learning Framework for Continuous Authentication in Zero-Trust Networks

Jie Ji, Shi Qiu, Siyu Ye, Xin LiuFuture 2026 · DOI

TL;DR

This paper introduces a novel Federated Incremental Learning framework (FIL-ZTA) for continuous authentication in Zero-Trust Networks. It tackles the challenges of evolving threats and data privacy by combining federated learning with incremental learning, achieving a significant reduction in error rates and improved detection of novel attacks while maintaining privacy.

What Problem Does This Paper Address?

Modern security architectures, especially Zero-Trust Networks (ZTA), demand continuous and adaptive identity authentication. Traditional authentication methods are insufficient for dynamic environments where user behavior and attack patterns constantly change (concept drift). Existing federated learning (FL) approaches struggle with this concept drift and non-independent and identically distributed (non-IID) data from different clients, leading to degraded model performance and susceptibility to catastrophic forgetting. Furthermore, data privacy concerns prevent the centralization of sensitive user data, creating data silos that hinder effective model training. This paper aims to bridge this gap by proposing a framework that can learn continuously from distributed data, adapt to evolving threats, and preserve user privacy.

Methods

The core of the proposed solution is a hybrid framework integrating Federated Learning (FL) with Incremental Learning (IL). Key components include:

  • Federated Incremental Learning (FIL-ZTA): This framework allows clients to train models locally on their incremental data and send updates to a central server for aggregation, all while preserving data privacy.
  • Dynamic Weighted Federated Aggregation (DWFA): This novel algorithm is central to mitigating concept drift. It dynamically adjusts aggregation weights in real-time based on client update quality and consistency. This ensures that the global model prioritizes recent and relevant information from clients, adapting to changing data distributions.
  • Dynamic Client Screening: Before aggregation, clients are screened based on data freshness, reputation, and update quality. This helps filter out potentially malicious or low-quality updates, improving the robustness of the global model.
  • Contribution-Aware Weighting: Clients' contributions are weighted based on their update's effectiveness (quality factor) and alignment with the consensus (consistency factor). This ensures that valuable updates have a greater impact on the global model.
  • Differential Privacy: To further enhance privacy, norm clipping and Gaussian noise are applied to client updates before aggregation, providing formal privacy guarantees.
  • Lightweight Local Authentication Model: A compact neural network architecture is designed for edge devices, balancing computational efficiency with effectiveness.

Datasets

The experiments were conducted using a combination of publicly available datasets and a custom user simulation system:

  • CIC-IDS-2018 dataset: Provided network attack traffic characteristics.
  • CIC-FlowMeter-V3 tool: Generated 78-dimensional network flow statistics.
  • Custom User Simulation System: Created 32-dimensional discrete identity authentication attributes (e.g., login time, device fingerprint, geographic location).

The combined dataset comprised 850,000 records, with 65% normal traffic and 35% attack traffic, covering six attack categories including brute force, DDoS, and web application attacks. To simulate concept drift, two new attack signatures were introduced at specific training steps.

Key Insights and Ideas

  • The Power of Hybrid Learning: Combining FL and IL is crucial for adaptive authentication in dynamic environments. FL handles privacy and distributed data, while IL enables continuous learning and adaptation to concept drift.
  • Dynamic Aggregation is Key: Simply averaging updates (like in FedAvg) is insufficient. Dynamically weighting client contributions based on their quality and consistency is essential for robust and efficient model updates.
  • Client Screening is Not Just for Security: Beyond preventing malicious attacks, dynamic client screening helps ensure that only relevant and up-to-date information influences the global model, directly combating concept drift.
  • Privacy Doesn't Have to Cripple Performance: The proposed differential privacy mechanism strikes a favorable trade-off, offering strong privacy protection with only a minor impact on model performance.
  • Edge-Friendly Design Matters: The lightweight local model architecture is critical for deployment on resource-constrained edge devices, enabling real-time authentication.

Key Results

  • Improved Authentication Accuracy: FIL-ZTA achieved an equal error rate (EER) of 2.15%, a 56.0% reduction compared to FedAvg.
  • Enhanced Detection of Novel Attacks: Detection rates for new attack types (D and E) were significantly higher, with FIL-ZTA achieving 17.8% and 19.9% higher detection rates than FedAvg, respectively.
  • Faster Adaptation to Concept Drift: FIL-ZTA recovered from concept drift in just two rounds, significantly faster than FedAvg (eight rounds).
  • Efficient Communication: Cumulative communication overhead was reduced by 47.4% compared to FedAvg.
  • Robustness Against Attacks: FIL-ZTA showed remarkable resilience to malicious clients, with only a 4.75% increase in EER at 30% malicious clients, compared to FedAvg's 18.92% increase.
  • Scalability: The asynchronous aggregation mechanism demonstrated good scalability, maintaining performance with an increasing number of clients.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Computational Complexity: The single-round training time for FIL-ZTA is slightly higher than FedAvg, which could impact real-time performance on highly resource-constrained devices.
  • Dynamic Evaluation Bottleneck: The dynamic evaluation mechanism might become a bottleneck with a very high number of clients.
  • Privacy Budget Refinement: While the privacy mechanism is effective, further refinement is needed for high-precision applications when dealing with extreme privacy settings.
  • Improved Dynamic Screening: The dynamic screening mechanism could be enhanced to better identify coordinated attack patterns.
  • Enhanced Drift Detection: Current drift detection relies on performance mutations, which might not be sensitive enough for progressive drift. More sensitive measures are needed.
  • Personalized Adaptation: Global aggregation can sometimes hinder flexibility for client-specific patterns. Future work could explore better ways to balance global learning with local personalization.
  • Lightweight Contribution Assessment: Investigating more lightweight measures for contribution assessment is suggested.
  • Adaptive Privacy Budgets: Exploring adaptive privacy budgets could further optimize the privacy-performance trade-off.
  • Architectural Innovations: Incorporating blockchain technology for enhanced security against attacks and exploring hierarchical federated structures for better global-local balance are potential avenues.

Summary updated

Abstract

Zero-trust architecture (ZTA) requires continuous and adaptive identity authentication to maintain security in dynamic environments. However, current federated learning (FL)-based authentication models often struggle to incorporate evolving attack patterns without experiencing catastrophic forgetting. Moreover, non-independent and identically distributed (non-IID) client data and concept drift frequently lead to degraded model robustness and personalization. To address these issues, this paper presents a hybrid learning framework that integrates federated learning with incremental learning (IL) for sustainable authentication. A Dynamic Weighted Federated Aggregation (DWFA) algorithm is developed to mitigate concept drift by adjusting aggregation weights in real time, ensuring that the global model adapts to changing data distributions. This approach enables continuous learning from distributed threat data while maintaining privacy and eliminating the need for historical data retention. Experimental results on real-world traffic datasets indicate that the proposed framework outperforms conventional FL baselines, reducing the overall error rate by approximately 56% and improving the detection rate for novel attack types by over 17.8%. Furthermore, the framework remains stable against performance decay while maintaining efficient communication overhead. This study provides an adaptive, privacy-preserving solution for identity authentication in zero-trust systems.
BibTeX
@inproceedings{ji2026hybrid,
  author = {Jie Ji and Shi Qiu and Siyu Ye and Xin Liu},
  title = {{A Hybrid Federated–Incremental Learning Framework for Continuous Authentication in Zero-Trust Networks}},
  booktitle = {Future},
  year = {2026},
  doi = {10.3390/fi18030154},
}
Figure