Trial and Trust: Addressing Byzantine Attacks with Comprehensive Defense Strategy

Gleb Molodtsov, Daniil Medyakov, Sergey Skorik, Nikolas Khachaturov, Shahane Tigranyan, Vladimir Aletov, Aram Avetisyan, Martin Takác, Aleksandr Beznosikov

Abstract

Recent advancements in machine learning have improved performance while also increasing computational demands. While federated and distributed setups address these issues, their structure is vulnerable to malicious influences. In this paper, we address a specific threat, Byzantine attacks, where compromised clients inject adversarial updates to derail global convergence. We combine the trust scores concept with trial function methodology to dynamically filter outliers. Our methods address the critical limitations of previous approaches, allowing functionality even when Byzantine nodes are in the majority. Moreover, our algorithms adapt to widely used scaled methods like Adam and RMSProp, as well as practical scenarios, including local training and partial participation. We validate the robustness of our methods by conducting extensive experiments on both synthetic and real ECG data collected from medical institutions. Furthermore, we provide a broad theoretical analysis of our algorithms and their extensions to aforementioned practical setups. The convergence guarantees of our methods are comparable to those of classical algorithms developed without Byzantine interference.

Introduction

In distributed and federated learning environments, the standard assumption is that participants contribute honest updates. However, "Byzantine" nodes—workers compromised by hardware failure or malicious intent—can transmit corrupted gradients that significantly impede or entirely prevent model convergence. Standard optimization algorithms like SGD or ADAM are notoriously fragile in these settings; a single poisoned update can cause the global model to diverge.

Existing defenses often rely on robust aggregation techniques, such as coordinate-wise medians. However, these typically require a majority of workers to be honest ($>50\%$), a condition often violated in real-world federated scenarios involving mobile devices or diverse medical institutions. Furthermore, adaptive attacks like ALIE (A Little Is Enough) or IPM (Inner Product Manipulation) can bypass simple filters by mimicking honest behavior.

Methodology: The Bant Framework

We introduce Bant (Byzantine Antidote via Trial Function and Trust Scores), a framework designed to maintain convergence even when malicious nodes outnumber honest ones. The core innovation is the Trial Function methodology, where the server maintains a small, clean validation dataset to evaluate worker updates.

Algorithm Variants

  1. Bant: This approach calculates contribution coefficients for each worker. If a worker's gradient reduces the loss on the trial set, it receives a positive score. A key feature is the use of weight momentum, which prevents the immediate silencing of honest workers due to stochastic noise, ensuring stability.
  2. AutoBant: This variant treats weight assignment as an auxiliary optimization problem, utilizing Mirror Descent to find the optimal weighting of updates that minimizes trial loss. This provides rigorous theoretical guarantees for non-convex optimization.
  3. SimBant: Specifically designed for classification, this variant measures the similarity between the logits of the server and workers to determine trustworthiness.

Bant is uniquely adaptable, supporting industry-standard scaled optimizers (ADAM/RMSProp), Local SGD, and partial client participation.

Experimental Evaluation

We validated the Bant framework across three domains: Computer Vision (CIFAR-10), Healthcare (proprietary ECG data for AFIB detection), and Recommender Systems (WEB30K).

Resilience to Sophisticated Attacks

On the CIFAR-10 dataset, we tested resilience against various attacks, including Label Flipping, Sign Flipping, and the adaptive ALIE attack.

Figure 1: Test accuracy of ResNet18 on CIFAR-10 under various Byzantine attacks (40-70% malicious nodes). Bant and its variants consistently maintain high accuracy where baselines fail.

As shown in Figure 1, under a 40% ALIE attack, Bant maintains approximately 85% accuracy, whereas standard ADAM and other defenses drop to near-random levels (~10%).

Performance on Heterogeneous and Medical Data

To simulate realistic federated learning, we evaluated performance on heterogeneously distributed data using a Dirichlet distribution.

Figure 2: Performance (NDCG@5) on Dirichlet-distributed data under Label Flipping and Random Gradient attacks. Bant variants demonstrate superior stability and ranking performance.

In the healthcare domain, Bant achieved a G-mean of 0.953 under a 60% Label Flipping attack on ECG data, significantly outperforming the baseline ADAM score of 0.262.

Key Insights

  • Breaking the Majority Rule: By anchoring trust in a server-side "ground truth" (the trial set), Bant functions as long as at least one honest worker exists.
  • Weight Momentum: This mechanism balances the need to filter outliers with the inherent noise of stochastic training, preventing the exclusion of honest but noisy participants.
  • Theoretical Optimality: Convergence rates for Bant are comparable to classical SGD, ensuring that security does not come at the cost of significant computational delay.

Conclusion

Bant provides a robust antidote to Byzantine interference in distributed learning. While it requires a small representative dataset on the server—which may pose challenges in extreme privacy-sensitive cases—its ability to handle majority-malicious networks and integrate with modern optimizers makes it a significant advancement. Future work will explore the use of synthetic data for the trial set to further mitigate privacy concerns.

Summary updated

Abstract

Recent advancements in machine learning have improved performance while also increasing computational demands. While federated and distributed setups address these issues, their structure is vulnerable to malicious influences. In this paper, we address a specific threat, Byzantine attacks, where compromised clients inject adversarial updates to derail global convergence. We combine the trust scores concept with trial function methodology to dynamically filter outliers. Our methods address the critical limitations of previous approaches, allowing functionality even when Byzantine nodes are in the majority. Moreover, our algorithms adapt to widely used scaled methods like Adam and RMSProp, as well as practical scenarios, including local training and partial participation. We validate the robustness of our methods by conducting extensive experiments on both synthetic and real ECG data collected from medical institutions. Furthermore, we provide a broad theoretical analysis of our algorithms and their extensions to aforementioned practical setups. The convergence guarantees of our methods are comparable to those of classical algorithms developed without Byzantine interference.
BibTeX
@inproceedings{molodtsov2026trial,
  author = {Gleb Molodtsov and Daniil Medyakov and Sergey Skorik and Nikolas Khachaturov and Shahane Tigranyan and Vladimir Aletov and Aram Avetisyan and Martin Takác and Aleksandr Beznosikov},
  title = {{Trial and Trust: Addressing Byzantine Attacks with Comprehensive Defense Strategy}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i29.39625},
}
Figure