BEV-SGD: Best Effort Voting SGD for Analog Aggregation Based Federated Learning against Byzantine Attackers

Xin Fan, Yue Wang, Yan Huo, Zhi Tian

Abstract

Analog Federated Learning Over the Air (FLOA) is a promising technology for high-efficiency distributed learning in edge computing. By allowing workers to transmit updates simultaneously over shared resources, FLOA achieves rapid aggregation but remains highly vulnerable to Byzantine attacks. Because the Parameter Server (PS) only receives a superimposed sum of signals, it cannot inspect individual updates to filter malicious data. This paper introduces BEV-SGD (Best Effort Voting Stochastic Gradient Descent), a novel power control policy where workers transmit at maximum power. Unlike the standard Channel Inversion (CI) method, BEV-SGD ensures that the collective signal of honest workers outweighs malicious interference, providing a robust defense in environments where traditional digital screening fails.

The Blind Spot of Analog Aggregation

Federated Learning (FL) typically involves devices sending digital packets to a server. In Federated Learning Over the Air (FLOA), devices transmit analog signals simultaneously. The physics of wireless superposition automatically "adds up" these signals, allowing the PS to receive the average gradient in a single step.

While efficient, this process creates a security blind spot. In digital FL, servers can use algorithms like the geometric median to discard outliers. In FLOA, the server is blind to individual contributions. A single attacker with a strong signal can "hijack" the global model, causing it to diverge. This paper analyzes the standard Channel Inversion (CI) mechanism and proposes Best Effort Voting (BEV) as a resilient alternative.

Methodology: CI vs. BEV

Channel Inversion (CI)

CI is the traditional "polite" approach to power control. Workers adjust their transmit power so that their signals arrive at the PS with identical amplitudes. While this ensures unbiased aggregation in benign settings, it makes the system fragile. By forcing honest workers with good channels to reduce their power to match those with poor channels, CI lowers the "collective volume" of the honest group, allowing a single Byzantine attacker to easily overwhelm the leveled signals.

Best Effort Voting (BEV)

BEV-SGD adopts a "shouting" approach. Every honest worker transmits their local gradient using their maximum transmit power. This treats aggregation like a noisy vote: if the majority of workers are honest and transmitting at max volume, malicious updates are treated as background noise rather than dominant signals. The authors prove that BEV can defend against attacks even if nearly half the workers are malicious, provided the workers are roughly isomorphic.

Figure 1: Training loss and test accuracy in a benign scenario (no attackers). CI and BEV perform comparably, with CI showing a slight edge in precision.

Experimental Evaluation

The researchers evaluated BEV-SGD using the MNIST dataset of handwritten digits on a Multilayer Perceptron (MLP) model. The setup involved 10 workers with varying numbers of Byzantine attackers. The attackers were modeled as "worst-case," sending the exact negative of the true gradient at maximum power.

Key Results

  • Robustness to Attackers: In simulations with multiple attackers, the CI method hit a "breaking point" rapidly. As shown in Figure 5, with 4 attackers, CI accuracy plummeted to approximately 10%, effectively failing to learn. In contrast, BEV-SGD maintained functional accuracy, reaching over 65% even under heavy attack.
  • The Benign Trade-off: In a perfectly clean environment, CI achieved slightly higher accuracy (~80% vs ~78% for BEV). This is because BEV introduces small amounts of "scaling noise" due to varying arrival strengths. However, this minor loss in precision is a necessary trade-off for the massive gain in Byzantine resilience.
  • Graceful Degradation: While CI's performance is binary (it either works or fails completely), BEV-SGD degrades gracefully as the number of attackers increases or the attacker's channel strength grows.
Figure 5: Training loss and test accuracy under varying numbers of Byzantine attackers. CI fails completely with 4 attackers, while BEV-SGD maintains stability.

Conclusion

BEV-SGD represents a significant step toward secure analog federated learning. By leveraging maximum transmit power as a form of "voting," it overcomes the inherent blindness of over-the-air aggregation. While CI remains preferable for guaranteed safe environments, BEV-SGD is essential for practical edge computing deployments where the honesty of every participating device cannot be guaranteed.

Summary updated

Abstract

As a promising distributed learning technology, analog aggregation based federated learning over the air (FLOA) provides high communication efficiency and privacy provisioning under the edge computing paradigm. When all edge devices (workers) simultaneously upload their local updates to the parameter server (PS) through commonly shared time-frequency resources, the PS obtains the averaged update only rather than the individual local ones. While such a concurrent transmission and aggregation scheme reduces the latency and communication costs, it unfortunately renders FLOA vulnerable to Byzantine attacks. Aiming at Byzantine-resilient FLOA, this paper starts from analyzing the channel inversion (CI) mechanism that is widely used for power control in FLOA. Our theoretical analysis indicates that although CI can achieve good learning performance in the benign scenarios, it fails to work well with limited defensive capability against Byzantine attacks. Then, we propose a novel scheme called the best effort voting (BEV) power control policy that is integrated with stochastic gradient descent (SGD). Our BEV-SGD enhances the robustness of FLOA to Byzantine attacks, by allowing all the workers to send their local updates at their maximum transmit power. Under worst-case attacks, we derive the expected convergence rates of FLOA with CI and BEV power control policies, respectively. The rate comparison reveals that our BEV-SGD outperforms its counterpart with CI in terms of better convergence behavior, which is verified by experimental simulations.
BibTeX
@article{fan2022bev,
  author = {Xin Fan and Yue Wang and Yan Huo and Zhi Tian},
  title = {{BEV-SGD: Best Effort Voting SGD for Analog Aggregation Based Federated Learning against Byzantine Attackers}},
  journal = {IEEE Internet of Things Journal},
  year = {2022},
  doi = {10.1109/jiot.2022.3164339},
}
Figure