DPFNAS: Differential Privacy-Enhanced Federated Neural Architecture Search for 6G Edge Intelligence

Yang Lv, Jin Cao, Ben Niu, Zhe Sun, Fengwei Wang, Fenghua Li, Hui LiarXiv 2025 · DOI · arXiv

As we transition toward 6G networks, "edge intelligence"—the deployment of AI directly onto decentralized devices like smartphones, drones, and IoT sensors—has become a primary objective. While Federated Learning (FL) offers a path to collaborative training without raw data sharing, it faces two critical bottlenecks: extreme data heterogeneity across devices and vulnerability to privacy-leakage attacks.

DPFNAS (Differential Privacy-Enhanced Federated Neural Architecture Search) is a novel framework designed to address these challenges by combining automated architecture search with a unique representation-based training strategy that protects privacy while drastically reducing communication overhead.

The Dual Challenges of Edge Intelligence

Standard Federated Learning often struggles in 6G environments for two reasons:

  1. Data Heterogeneity: A single global model rarely performs well across diverse local datasets (e.g., a sensor in a factory vs. a smartphone in a city).
  2. Privacy Risks: Even without sharing raw data, malicious actors can use "gradient leakage" to reconstruct private information from shared model updates.

Existing solutions often rely on "supernets"—massive, predefined structures from which clients sample sub-models. However, these are frequently too computationally heavy for edge devices and lack robust, automated privacy guarantees.

The DPFNAS Framework

The proposed DPFNAS framework operates in two distinct phases: Privacy-Aware NAS (PANAS) and DP-Enhanced Representation-Based Federated Training (DPRBFT).

The two-phase DPFNAS framework: Phase 1 (PANAS) optimizes the architecture, while Phase 2 (DPRBFT) handles collaborative training via feature sharing.

Phase 1: Privacy-Aware NAS (PANAS)

Instead of enforcing a uniform model, PANAS allows each device to search for its own optimal architecture using a Genetic Algorithm (GA). To ensure these models remain lightweight, the search space is "block-based," utilizing efficient modules like Depthwise Separable Convolutions.

Crucially, the framework employs Bayesian Optimization (BO) to automatically tune hyperparameters—such as learning rates and noise scales—to satisfy each device's specific Differential Privacy (DP) budget. This ensures that the privacy-utility trade-off is optimized automatically for every participant.

Phase 2: DP-Enhanced Representation-Based Federated Training (DPRBFT)

Once architectures are established, DPFNAS departs from traditional FL by sharing intermediate feature representations ($z_k$) rather than full model weights.

The DPRBFT workflow: Clients apply personalized DP noise to intermediate representations before uploading them to the server for global head training.

In this workflow:

  • Feature Sharing: Clients process data through their local "backbone" and send the resulting features to the server.
  • Personalized DP: Before transmission, clients apply DP noise to these features, preventing the server from reconstructing the original input images.
  • Global Head Training: The server trains a shared "global head" model using these protected features and distributes the updated head parameters back to the clients.

Performance and Security Analysis

The framework was evaluated on CIFAR-10 and CIFAR-100 datasets under non-IID (Independent and Identically Distributed) conditions to simulate real-world edge bias.

Key Results:

  • Superior Accuracy: On the complex CIFAR-100 benchmark, DPFNAS achieved 47.28% accuracy, outperforming the state-of-the-art PerFedRLNAS baseline by 6.82%.
  • Communication Efficiency: By sharing representations instead of millions of model parameters, DPFNAS reduced the communication cost required to reach target accuracy by 20x compared to traditional methods.
  • Robust Privacy: In "Feature Space Hijacking" tests, DPFNAS effectively neutralized reconstruction attacks. With a strict privacy budget ($\epsilon = 5.0$), reconstructed images were rendered completely unrecognizable (MSE $\geq 0.67$).

Conclusion and Future Directions

DPFNAS demonstrates that privacy and efficiency do not have to be mutually exclusive in 6G edge intelligence. By moving from parameter-averaging to representation-sharing and automating the privacy-utility trade-off, the framework provides a scalable solution for heterogeneous edge environments. Future work will focus on optimizing for system heterogeneity (e.g., varying battery lives) and extending the framework to natural language processing and time-series IoT applications.

Summary updated

Abstract

The Sixth-Generation (6G) network envisions pervasive artificial intelligence (AI) as a core goal, enabled by edge intelligence through on-device data utilization. To realize this vision, federated learning (FL) has emerged as a key paradigm for collaborative training across edge devices. However, the sensitivity and heterogeneity of edge data pose key challenges to FL: parameter sharing risks data reconstruction, and a unified global model struggles to adapt to diverse local distributions. In this paper, we propose a novel federated learning framework that integrates personalized differential privacy (DP) and adaptive model design. To protect training data, we leverage sample-level representations for knowledge sharing and apply a personalized DP strategy to resist reconstruction attacks. To ensure distribution-aware adaptation under privacy constraints, we develop a privacy-aware neural architecture search (NAS) algorithm that generates locally customized architectures and hyperparameters. To the best of our knowledge, this is the first personalized DP solution tailored for representation-based FL with theoretical convergence guarantees. Our scheme achieves strong privacy guarantees for training data while significantly outperforming state-of-the-art methods in model performance. Experiments on benchmark datasets such as CIFAR-10 and CIFAR-100 demonstrate that our scheme improves accuracy by 6.82\% over the federated NAS method PerFedRLNAS, while reducing model size to 1/10 and communication cost to 1/20.
BibTeX
@misc{lv2025dpfnas,
  author = {Yang Lv and Jin Cao and Ben Niu and Zhe Sun and Fengwei Wang and Fenghua Li and Hui Li},
  title = {{DPFNAS: Differential Privacy-Enhanced Federated Neural Architecture Search for 6G Edge Intelligence}},
  year = {2025},
  eprint = {2509.23030},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.23030},
}
Figure