In an effort to protect distributed vehicular networks from cyber threats while preserving communication efficiency, engineering researchers have developed a decentralized federated learning algorithm capable of operating reliably under Byzantine adversarial attacks without requiring exact mathematical gradient information. Modern intelligent transportation systems increasingly rely on the Internet of Vehicles, an architectural paradigm that interconnects onboard sensors, vehicle controllers, and roadside units to elevate navigational safety and coordinate real-time traffic management. However, vehicular networks typically encounter distributed data silos and severe operational vulnerabilities, as sharing raw sensor readings across wireless links introduces profound security and privacy risks. While decentralized federated learning offers a path forward by enabling telematics clients to train models locally and exchange only parameter updates, real-world deployments face severe hazards from compromised nodes that deliberately broadcast erroneous or conflicting data to derail the global optimization process. To address these vulnerabilities simultaneously, the research team designed a Byzantine-resilient decentralized bandit online federated learning framework that successfully curbs malicious disruptions and reduces optimization query costs.
The study investigates the fundamental tension between network decentralization, computational constraints, and security vulnerabilities within dynamic vehicular settings. In conventional centralized federated learning, a single aggregation server collects parameter updates from distributed clients, constructs an updated global model, and redistributes that model back across the network. Although centralized aggregation avoids exposing raw local data, it introduces a dangerous single point of failure and demands substantial resource overhead that can overwhelm wireless channels in dense traffic corridors. Conversely, decentralized federated learning eliminates the central coordinator entirely by organizing participating vehicle clients into peer-to-peer communication graphs. In these decentralized structures, each regular vehicle interacts exclusively with immediate topological neighbors, executing local model parameter updates and collaborative aggregation algorithms. However, removing the centralized authority makes the system acutely vulnerable to malicious nodes that exploit peer-to-peer trust. Without a central clearinghouse to inspect and sanitize incoming data, adversarial perturbations can rapidly propagate across local neighborhoods, corrupting consensus and causing client trajectories to diverge disastrously.
Compounding the security challenge is the physical reality that onboard vehicular sensors and telematics controllers often cannot compute exact analytical gradients. Most classical distributed optimization schemes assume that accurate first-order derivative data is continually accessible during local training routines. In practical vehicular environments, however, onboard microprocessors encounter severe hardware limitations, power caps, and numerical instabilities such as gradient vanishing or exploding across deep neural layers. To bypass explicit gradient evaluations, the researchers turned to bandit feedback techniques, where distributed agents interact with loss functions solely by evaluating objective values at specific sampling points rather than calculating analytical partial derivatives. Multi-agent distributed online optimization historically deployed either one-point bandit feedback or two-point bandit feedback.
To overcome the sampling bottlenecks inherent to traditional zeroth-order approaches, the proposed framework integrates a specialized One-Point Residual Feedback mechanism into the local optimization loop. Conventional two-point gradient estimators require clients to query their objective functions twice during every communication round, effectively doubling the evaluation overhead imposed on low-power vehicular computational units. In contrast, the one-point residual approach constructs an unbiased gradient estimate by evaluating only the current perturbed objective state and subtracting the residual function value stored directly from the previous iteration. By reusing historical feedback, the algorithm requires just one fresh objective-function evaluation per client per round after initial startup, slashing the dominant oracle-query overhead by fifty percent while retaining the estimation precision and asymptotic scaling advantages typical of two-point schemes.
To defend against malicious agents attempting to manipulate the learning process, the framework incorporates an explicit B-local Byzantine adversary model coupled with structural graph-robustness conditions. In distributed network theory, a Byzantine attack represents one of the most covert and destructive threat categories, as compromised nodes are permitted to deviate arbitrarily from standard execution protocols, including transmitting completely contradictory values to different topological neighbors in the same communication round. Under the adopted model, the network topology is assumed to form a (2B+1)-robust undirected communication graph, while the adversary set remains B-local, meaning that no regular client has more than B Byzantine neighbors within its immediate communication vicinity. These connectivity criteria guarantee that the subset of honest, non-faulty clients remains sufficiently interconnected and populous to prevent malicious nodes from isolating benign vehicles or dominating neighborhood consensus through coordinated numerical deception.
The defensive mechanism operates locally at each regular vehicle client without requiring global network awareness or external administrative intervention. During the local communication round, each regular node collects parameter states broadcast by its immediate neighbors and arranges them in sorted numerical order. Under the resilient filtering mechanism, the client identifies and discards the B largest and B smallest parameter values from the collected neighborhood set, establishing a trimmed neighborhood subset. If fewer than B parameter values exceed or fall below the client’s own local state, all such extreme values are discarded. This trimming process ensures that extreme numerical outliers injected by Byzantine adversaries are eliminated before local aggregation takes place. Even if an adversarial value falls within the untrimmed range, it inherently lies closer to legitimate regular states than the excluded values, thereby neutralizing its disruptive influence. Following localized outlier filtering, the client updates its model parameter using a projected gradient step that constrains updates within a bounded convex feasible set.
The authors provide rigorous mathematical convergence guarantees for their algorithmic design under convex, Lipschitz-continuous local objective functions over a fixed communication topology. Evaluating optimization efficacy in dynamic vehicular environments requires measuring dynamic regret, which quantifies the cumulative difference between the client’s sequential decisions and the time-varying optimal parameter trajectory in hindsight. The formal convergence analysis establishes that the effective neighborhood mixing matrix, derived after pruning Byzantine neighbors, remains row-stochastic with strictly positive diagonal entries and uniform lower bounds on non-zero weights. Through backward products of these dynamic mixing matrices, the researchers demonstrated that the proposed algorithm attains sublinear dynamic regret, confirming that regular vehicle clients can asymptotically track the evolving optimal model trajectory despite the persistent injection of arbitrary data by Byzantine attackers.
To validate the theoretical framework empirically, the researchers conducted extensive simulation experiments using standard machine learning benchmarks alongside vehicular vision tasks. The static evaluation utilized the MNIST and Fashion-MNIST image classification datasets, each containing 70,000 grayscale images across 10 categories, divided into 54,000 training, 6,000 validation, and 10,000 testing samples. The experimental setup simulated an intelligent vehicular network comprising 12 distributed clients, structured with 10 regular clients and 2 Byzantine adversaries across 100 communication rounds. Onboard processing was modeled through convolutional neural networks featuring two convolutional layers, ReLU activations, max-pooling stages, and fully connected classification layers. The experimental framework subjected the network to two distinct non-adaptive threat models: sign-flipping attacks, where adversaries amplify and reverse their model updates to push parameters in opposing directions, and Gaussian attacks, where adversaries perturb locally trained weights with independent random noise before broadcasting them to peers.
Comparative evaluations against standard decentralized learning baselines demonstrated the resilience of the proposed architecture under aggressive Byzantine disruption. When benchmarked against decentralized federated averaging without defensive filtering, the proposed algorithm prevented catastrophic model divergence, improving final classification accuracy under Gaussian attacks by 62.05 percentage points on Fashion-MNIST and 64.66 percentage points on MNIST. When compared against Krum, a classical distance-based outlier filtering approach, the proposed framework maintained accuracy advantages of 8.56 and 6.34 percentage points across Fashion-MNIST and MNIST, respectively. Statistical analysis revealed a rank-biserial correlation of 1.000, underscoring a consistent performance advantage over both baseline schemes. Meanwhile, the algorithm achieved performance virtually identical to coordinate-wise Trimmed Mean, recording an absolute accuracy difference of merely 0.02 percentage points on both static datasets while operating within the decentralized zeroth-order bandit execution framework.
To isolate the precise contributions of the filtering mechanism and the gradient estimation structure, the authors executed component-level ablation experiments on a simulated 12-client dynamic vehicular network over 100 communication rounds using 10 matched random seeds. These empirical measurements confirmed that the residual mechanism halving function queries incurs no degradation in model accuracy or adversarial resilience.
Scalability assessments further confirmed that the algorithm sustains high learning performance across expanding network topologies. The research team evaluated system performance across simulated fleets consisting of 12, 30, 50, and 100 clients, with Byzantine node counts scaled proportionally to 2, 5, 8, and 16 adversaries to maintain a steady attack ratio. As network size expanded from 12 to 100 nodes, final model accuracy exhibited only a minor reduction from 98.95 percent down to 95.54 percent. While the number of communication rounds required to cross a 95 percent accuracy threshold increased from 6 rounds in the 12-client setup to 86 rounds in the 100-client system, the algorithm preserved stable convergence and robust classification throughout.
To evaluate performance under vehicular mobility, the researchers deployed the German Traffic Sign Recognition Benchmark across a dynamic, window-based topology-switching environment. The traffic sign dataset comprises 39,209 training images and 12,630 test images across 43 classes, processed by three-channel convolutional networks. Vehicular mobility was simulated over 100 communication rounds divided into five consecutive 20-round windows, with the underlying network graph switching to a new predefined topology at rounds 20, 40, 60, and 80. Although these periodic network reconfigurations induced brief accuracy fluctuations at switching boundaries due to sudden changes in neighborhood connectivity and Byzantine neighbor visibility, the algorithm recovered swiftly across subsequent rounds. Under Gaussian attacks, the resilient filtering mechanism preserved a stable upward learning trajectory throughout the topology changes, whereas non-resilient aggregation suffered severe accuracy collapse, illustrating the practical viability of localized filtering in mobile vehicular channels.
Despite these favorable findings, the researchers emphasized several technical limitations that define boundaries for future investigation. The formal theoretical proofs are established for scalar parameter spaces and convex, Lipschitz-continuous cost functions over static communication graphs, whereas practical deep neural networks operate over highly non-convex, high-dimensional parameter spaces. Additionally, the experimental evaluations focused on static, non-adaptive Byzantine attackers whose identities and attack behaviors remained invariant throughout training, leaving adaptive adversaries who strategically mask their perturbations or collude across rounds outside the current theoretical guarantees. Furthermore, because model parameters were transmitted in plaintext, the architecture remains vulnerable to wireless eavesdropping and passive traffic analysis. The authors noted that future developments must focus on integrating cryptographic defenses, designing adaptive filtering thresholds capable of defeating coordinated collusion, extending non-convex theoretical convergence bounds, and conducting hardware-in-the-loop trials in operational physical vehicular networks.
Subject of Research: Byzantine-resilient decentralized online federated learning with bandit feedback in vehicular networks
Article Title: Byzantine-resilient decentralized online federated learning with Bandit feedback for IoV networks
Article References: Guo, G., Zhao, L., Chen, S., Zhang, F., Fang, J., & Zhang, W. (2026). Byzantine-resilient decentralized online federated learning with Bandit feedback for IoV networks. PLOS One, 21(11), e0362324. https://doi.org/10.1371/journal.pone.0362324
Image Credits: AI Generated
DOI: 10.1371/journal.pone.0362324
Keywords: Internet of Vehicles, Decentralized Federated Learning, Byzantine Fault Tolerance, Bandit Feedback, One-Point Residual Feedback, Dynamic Regret, Network Security, Autonomous Driving, Adversarial Robustness, Intelligent Transportation Systems, Zeroth-Order Optimization, Traffic Sign Recognition, Peer-to-Peer Networks, Robust Aggregation, Distributed Optimization, Communication Topology, Machine Learning Security
Cite Scienmag News
Veronica Carney. (November 21, 2026). Decentralized Federated Learning Algorithm Shields Connected Vehicles From Byzantine Attacks. Scienmag. https://scienmag.com/decentralized-federated-learning-algorithm-shields-connected-vehicles-from-byzantine-attacks/
Veronica Carney. "Decentralized Federated Learning Algorithm Shields Connected Vehicles From Byzantine Attacks." Scienmag, 21 November 2026, https://scienmag.com/decentralized-federated-learning-algorithm-shields-connected-vehicles-from-byzantine-attacks/. Accessed 21 November 2026.
Veronica Carney. "Decentralized Federated Learning Algorithm Shields Connected Vehicles From Byzantine Attacks." Scienmag. November 21, 2026. https://scienmag.com/decentralized-federated-learning-algorithm-shields-connected-vehicles-from-byzantine-attacks/








