Abstract
Over-the-Air (OtA) computation is a promising technique for efficient data aggregation in massive IoT and Federated Learning. However, its uncoded analog nature makes it vulnerable to eavesdropping, especially by cooperative attackers. This paper proposes a security framework using correlated "zero-forced" artificial noise (AN). By designing AN that cancels out at the legitimate server while interfering with eavesdroppers, the system achieves high security without sacrificing computation accuracy. Simulations demonstrate that the proposed design effectively mitigates the threat of cooperative eavesdropping in complex-valued fading channels.
Introduction
In massive Internet of Things (IoT) and Federated Learning (FL) deployments, aggregating data from numerous devices often leads to significant communication bottlenecks. Over-the-Air (OtA) computation addresses this by allowing devices to transmit simultaneously, utilizing the natural superposition property of the wireless channel to perform summation. While efficient, OtA is inherently vulnerable to security threats because it typically relies on uncoded analog transmissions.
A critical threat identified in this work is the presence of cooperative eavesdroppers. While individual attackers might be hindered by random channel phase misalignments, a group of cooperating eavesdroppers can pool their observations to reconstruct the aggregated data accurately. This paper introduces a novel estimation-theoretic security framework to protect OtA computation against such sophisticated adversaries.
Proposed Methodology
The core of the proposed solution is the use of Correlated Artificial Noise (AN). Unlike standard noise injection, which can degrade the accuracy of the intended computation at the server, the authors employ a Zero-Forcing (ZF) design.
Zero-Forced Artificial Noise
In this approach, users transmit carefully calibrated noise signals alongside their data. These noise components are designed such that their sum at the legitimate server is exactly zero, ensuring no "self-interference" or degradation of the computation result. To achieve this, the system distributes the "responsibility" of noise cancellation among users, often leveraging those with superior channel conditions to act as noise-cancelers.
Security Framework
The authors treat the group of $L$ cooperative eavesdroppers as a single "virtual eavesdropper" with multiple distributed antennas. The security objective is formulated as a max-min optimization problem: maximizing the minimum Mean Squared Error (MSE) at the eavesdroppers while minimizing the approximation error at the server. This framework accounts for the spatial diversity and joint estimation capabilities of the attackers.
Performance Evaluation
The proposed framework was evaluated using synthetic simulations involving 10 users and up to 14 eavesdroppers in a complex-valued block fading environment.
Key Findings
- Impact of Cooperation: There is a significant "security gap" between single and cooperative eavesdroppers. Cooperation allows attackers to exploit spatial diversity to overcome the natural protection provided by phase misalignment in complex channels.
- Effectiveness of ZF-AN: The optimized ZF-AN design significantly reduces the advantage of cooperation. As shown in the results, the system maintains a high security level (normalized MSE close to 0.9) even as the number of eavesdroppers increases, while keeping the server's approximation error low.
- Channel Characteristics: The inherent security of complex-valued channels is notably higher than that of real-valued models, as random phase shifts naturally degrade the attacker's ability to sum signals correctly.
Conclusion
This research demonstrates that OtA computation can be made resilient to cooperative eavesdropping through the strategic use of correlated artificial noise. By ensuring the noise cancels out at the intended receiver, the system achieves a robust trade-off between security and computational accuracy. Future work will address the impact of imperfect Channel State Information (CSI) and explore joint user selection strategies to further optimize performance.
Summary updated