Client-Verifiable and Efficient Federated Unlearning in Low-Altitude Wireless Networks

Yuhua Xu, Mingtao Jiang, Chenfei Hu, Yinglong Wang, Chuan Zhang, Meng Li, Ming Lu, Liehuang ZhuarXiv 2026 · arXiv

Here is a concise, engaging summary of the paper "Client-Verifiable and Efficient Federated Unlearning in Low-Altitude Wireless Networks."

TL;DR

VerFU introduces a privacy-preserving and client-verifiable federated unlearning framework tailored for highly dynamic low-altitude wireless networks (LAWNs). By combining linear homomorphic hashes, commitment schemes, and homomorphic encryption, it allows mobile edge devices to independently verify that a UAV-assisted server has actually erased their historical data contributions without needing access to original data samples.

What Problem Does This Paper Address?

In Low-Altitude Wireless Networks (LAWNs), unmanned aerial vehicles (UAVs) act as servers to coordinate Federated Learning (FL) among mobile Integrated Sensing and Communication (ISAC) devices (like smart vehicles or cameras). Because of privacy regulations like GDPR and the "Right to Be Forgotten" (RTBF), these networks must support Federated Unlearning (FUL)—the ability to completely remove a departing device's historical data influence from the trained global model.

However, existing FUL frameworks blindly trust the server to honestly execute the unlearning process. In reality, a third-party UAV server might secretly retain a high-value client's gradients to keep model accuracy high, risking privacy leaks. Furthermore, existing verification methods either require clients to keep original data samples (which is impractical for streaming edge data) or process unlearning requests sequentially, creating a massive bottleneck when multiple devices constantly enter and leave the network.

Methods

The authors propose VerFU, a framework designed to achieve secure, parallel, and sample-free unlearning verification:

  • Privacy-Preserving Aggregation & Unlearning: VerFU uses Paillier Homomorphic Encryption (HE). This allows the UAV server to aggregate updates and subtract historical contributions directly within the encrypted domain, keeping raw gradients hidden.
  • Tamper-Proof Records: Before uploading updates, ISAC devices generate a Linear Homomorphic Hash (LHH) of their gradients and wrap it in a cryptographic commitment scheme. This creates an unalterable record of their historical contributions.
  • Client State Tagging: To handle highly dynamic networks, VerFU introduces a status indicator ($f_i = 1$ for active training, $f_i = -1$ for unlearning). This allows the server to process multiple concurrent unlearning requests in parallel rather than sequentially.
  • Sample-Free Verification: Utilizing the linear composability of LHH, unlearning clients can jointly verify if the decrypted global model accurately reflects the removal of their data by checking if the hash of the final model matches the combined linear hash of the updates.

The overall architecture and workflow of VerFU are illustrated below:

Fig. 1: The architecture of VerFU.

Datasets

The framework was evaluated using a simulated LAWN environment with 500 ISAC devices on three standard image classification benchmarks:

  • MNIST
  • Fashion-MNIST
  • EMNIST

To mimic real-world edge environments, the data was distributed non-IID across the devices using a Dirichlet distribution ($\alpha = 1$).

Key Insights and Ideas

  • Cryptographic vs. Data-Driven Verification: Instead of relying on data-driven verification (like injecting backdoors or monitoring loss changes on saved samples), VerFU shifts verification entirely to the cryptographic domain. This eliminates the need for edge devices to store historical raw data.
  • The Power of Linear Homomorphic Hashes: Because LHH allows mathematical operations on hashes to mirror operations on the underlying data, clients can verify complex server-side subtractions simply by multiplying and dividing the corresponding hashes.
  • Parallelized Efficiency: By batching unlearning requests via state tags, the framework prevents the network from grinding to a halt when multiple drones or vehicles disconnect simultaneously.

Key Results

  • High Model Utility: Even at a massive 40% unlearning rate (where nearly half the clients request data erasure), the global model maintains robust accuracy (98.33% on MNIST, 88.69% on Fashion-MNIST, and 84.86% on EMNIST). The model fully recovers from unlearning disruptions within just 2 to 4 normal training rounds.
  • Low Communication Overhead: VerFU requires roughly 74.42 MB of communication overhead per round, which is significantly lower than alternative privacy-preserving verification baselines like PoLHE (~853.82 MB).
  • Blazing Fast Verification: VerFU achieves the lowest verification latency among all tested baselines, taking only 11.53s to 12.06s to complete verification, compared to zkPoT (which takes up to 24,746s) and PoLHE (up to 191,565s).

Limitations and Future Work

The authors note that VerFU operates under the assumption that the ISAC devices themselves behave honestly during the verification phase. Future research directions include:

  1. Exploring scenarios where clients might be malicious, upload corrupted verification parameters, or collude to disrupt the unlearning process.
  2. Further optimizing the homomorphic encryption overhead to accommodate extremely resource-constrained, low-power IoT sensor nodes.

Summary updated

Abstract

In low-altitude wireless networks (LAWN), federated learning (FL) enables collaborative intelligence among unmanned aerial vehicles (UAVs) and integrated sensing and communication (ISAC) devices while keeping raw sensing data local. Due to the "right to be forgotten" requirements and the high mobility of ISAC devices that frequently enter or leave the coverage region of UAV-assisted servers, the influence of departing devices must be removed from trained models. This necessity motivates the adoption of federated unlearning (FUL) to eliminate historical device contributions from the global model in LAWN. However, existing FUL approaches implicitly assume that the UAV-assisted server executes unlearning operations honestly. Without client-verifiable guarantees, an untrusted server may retain residual device information, leading to potential privacy leakage and undermining trust. To address this issue, we propose VerFU, a privacy-preserving and client-verifiable federated unlearning framework designed for LAWN. It empowers ISAC devices to validate the server-side unlearning operations without relying on original data samples. By integrating linear homomorphic hash (LHH) with commitment schemes, VerFU constructs tamper-proof records of historical updates. ISAC devices ensure the integrity of unlearning results by verifying decommitment parameters and utilizing the linear composability of LHH to check whether the global model accurately removes their historical contributions. Furthermore, VerFU is capable of efficiently processing parallel unlearning requests and verification from multiple ISAC devices. Experimental results demonstrate that our framework efficiently preserves model utility post-unlearning while maintaining low communication and verification overhead.
BibTeX
@misc{xu2026client,
  author = {Yuhua Xu and Mingtao Jiang and Chenfei Hu and Yinglong Wang and Chuan Zhang and Meng Li and Ming Lu and Liehuang Zhu},
  title = {{Client-Verifiable and Efficient Federated Unlearning in Low-Altitude Wireless Networks}},
  year = {2026},
  eprint = {2603.29688},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.29688},
}
Figure