Abstract
Federated Learning (FL) is susceptible to Byzantine attacks where malicious clients submit poisoned updates to degrade the global model. We propose FedCC (Consensus Confirmation), a novel framework that introduces a validation step by a randomly selected "jury" of clients. This mechanism ensures that only updates verified to improve model performance are integrated, significantly enhancing robustness against data poisoning without sacrificing accuracy in benign environments.
Introduction
In standard Federated Learning, a central server aggregates local model updates from distributed clients. While efficient, this architecture is vulnerable to Byzantine attackers who may upload "poisoned" updates—trained on mislabeled data or containing noise—to corrupt the global model. Traditional defenses like Krum or Median aggregation often struggle with non-IID data or can be bypassed by sophisticated attacks. FedCC addresses this by treating model updates as hypotheses requiring peer validation before finalization.
Methodology: The Consensus Confirmation Framework
FedCC augments the standard FL workflow with a validation layer. In each round $r$, the process follows five key stages:
- Candidate Generation: The server selects $n$ clients to perform local training. Their updates are aggregated to form a candidate global weight ($\tilde{w}_r$).
- Consensus Selection: A separate, random set of $l$ clients is chosen to act as a validation "jury."
- Local Validation: These $l$ clients evaluate the candidate weight ($\tilde{w}_r$) against the previous round's weight ($w_{r-1}$) using their local datasets.
- Confirmation: Clients vote on whether the candidate represents a genuine improvement.
- Final Decision: The server adopts the candidate as the new global weight ($w_r$) only if a majority of the jury confirms it. Otherwise, the model remains at $w_{r-1}$.
This dual-group approach creates a probabilistic barrier: an attacker must not only poison the training group but also control the majority of the randomly selected validation group to succeed.
Experimental Setup
The framework was evaluated using the MNIST dataset in a non-IID configuration. Byzantine behavior was simulated via label-flipping attacks (e.g., mapping label $y$ to $9-y$). The performance of FedCC was compared against the standard FedAVG algorithm.
Results
Robustness to Poisoning
Mathematical analysis and simulations demonstrate that FedCC drastically reduces the probability of a successful attack compared to standard FL. While the risk in standard FL grows linearly with the ratio of attackers, FedCC's combinatorial validation step causes the success probability to drop sharply as the total number of clients increases.
Model Accuracy and Stability
Experimental results highlight FedCC's superior performance under adversarial conditions. In scenarios with 20% attackers, FedCC maintained stable accuracy near 90%, whereas FedAVG experienced significant volatility and performance degradation. Crucially, in the absence of attacks, FedCC performs on par with FedAVG, indicating that the validation step does not introduce a performance penalty during normal operation.
Discussion and Future Work
FedCC provides a robust defense by prioritizing model stability over rapid updates. By rejecting suspicious candidates, it prevents "catastrophic forgetting" induced by poisoned data. However, the additional communication round for validation may increase overhead in bandwidth-limited settings. Future research will explore FedCC's efficacy on more complex datasets like CIFAR-100 and investigate strategies to mitigate potential validator collusion in environments with extremely high attacker ratios.
Summary updated