Abstract
Federated Learning (FL) enables decentralized devices to collaboratively train machine learning models without sharing raw data. However, implementing FL over wireless channels introduces significant communication bottlenecks and privacy risks. This paper presents FLORAS (Federated Learning using ORthogonAl Sequences), a novel framework designed for over-the-air model aggregation. Unlike traditional methods, FLORAS eliminates the need for Channel State Information at the Transmitter (CSIT) and provides mathematically rigorous differential privacy (DP) guarantees. By leveraging the properties of orthogonal sequences and a projection-based receiver, FLORAS achieves superior accuracy in noisy environments while protecting both individual data points and client participation.
Introduction
As the number of edge devices grows, traditional "upload-then-aggregate" federated learning protocols face severe scalability issues due to limited wireless bandwidth. Over-the-Air Computation (AirComp) has emerged as a solution, allowing signals to overlap and sum naturally in the air. However, standard AirComp typically requires precise Channel State Information at the Transmitter (CSIT) to perform channel inversion. In practical wireless settings, obtaining accurate CSIT is power-intensive and prone to errors, particularly when a user is in a deep fade.
Furthermore, model updates themselves can leak sensitive information about the underlying training data. While differential privacy (DP) is often used to mitigate this, existing wireless FL solutions frequently lack rigorous guarantees or fail to protect the system at the "client-level" (ensuring an adversary cannot determine if a specific user participated in the training round).
The FLORAS Framework
FLORAS introduces a robust architecture for wireless FL that addresses both communication efficiency and privacy. The core innovation lies in its use of orthogonal spreading sequences, similar to Code Division Multiple Access (CDMA) technology.
CSIT-Free Over-the-Air Aggregation
In FLORAS, each client is assigned a unique orthogonal sequence from a predefined set. During the uplink phase, clients transmit their model updates modulated by these sequences. Crucially, clients do not need to "invert" their channel or know their specific channel gains. The Base Station (BS) handles the aggregation by using a "projector" constructed from estimated channel gains. This shift from transmitter-side compensation to receiver-side projection makes the system significantly more robust to fading and eliminates the overhead of CSIT acquisition.
Privacy through Redundancy and Anonymity
The framework ensures privacy through two primary mechanisms:
- Sequence Anonymity: While the BS knows the set of all possible orthogonal sequences, it does not know which specific sequence belongs to which user. This is achieved through a random permutation mechanism using a shared confidential key.
- Noise Injection via Projection: The BS projects the received signal onto the entire set of sequences, including a subset of "unused" sequences. This process naturally injects noise into the final aggregate.
Differential Privacy via Cauchy Noise
A significant theoretical contribution of the FLORAS paper is the discovery that the projection process transforms standard Gaussian channel noise into Cauchy-distributed noise. Because Cauchy distributions have "fat tails" compared to Gaussian distributions, they provide exceptionally strong differential privacy guarantees.
To ensure the model still converges, the authors implement a truncation operation. Since pure Cauchy noise has infinite variance—which would typically destabilize machine learning training—truncating the noise allows the system to maintain a bounded variance while preserving the rigorous DP properties. The authors provide proofs for both item-level DP (protecting individual data samples) and client-level DP (protecting user participation).
Experimental Evaluation
The researchers evaluated FLORAS using the MNIST dataset of handwritten digits across two distinct scenarios:
- IID (Independent and Identically Distributed): Data is distributed randomly across 20 clients.
- Non-IID: Data is sorted by labels, meaning each client only sees specific digits, simulating the data heterogeneity common in real-world edge deployments.
Performance and Accuracy
In low-power scenarios (0 dB SNR), FLORAS demonstrated remarkable resilience. It achieved approximately 7.5% higher test accuracy than standard channel-inversion AirComp on IID data. In the more challenging non-IID setting, the performance gap widened to 10.2%.
Convergence and Privacy Tradeoffs
Theoretical analysis confirmed that FLORAS maintains an $O(1/T)$ convergence rate, ensuring the model reliably improves over time despite the injected noise. Experiments regarding the privacy-accuracy tradeoff showed that even with high privacy settings (utilizing a large number of unused sequences), the loss in test accuracy was minimal—ranging between 2.5% and 3.5% compared to a non-private baseline.
Conclusion
FLORAS represents a significant step forward for private wireless federated learning. By moving the complexity of channel handling to the Base Station and utilizing the unique properties of orthogonal sequence projection, it provides a "CSIT-free" solution that is both efficient and private. Future work may explore extending this framework to Non-Orthogonal Multiple Access (NOMA) to further improve scalability for massive IoT networks and tightening the lower bounds of the differential privacy analysis.
Summary updated