Byzantine-Robust Aggregation
What you'll learn: How to protect federated learning systems from malicious or faulty clients that send corrupted model updates.
The Problem: Bad Actors in Federated Learning
In federated learning, you aggregate model updates from many clients. But what if some clients are malicious—intentionally sending poisoned gradients to sabotage the global model? Or simply faulty, producing corrupted updates due to hardware issues?
The term "Byzantine" comes from the Byzantine Generals Problem in distributed systems: how do you reach consensus when some participants might be traitors? In federated learning, a Byzantine client might send wildly incorrect gradients to:
- Inject backdoors into the model
- Degrade overall accuracy
- Bias predictions toward certain outcomes
Standard Federated Averaging blindly trusts all updates, making it vulnerable.
Byzantine-Robust Aggregation Methods
Instead of simple averaging, robust aggregation methods detect and mitigate malicious updates:
- Coordinate-wise median: For each parameter, take the median across all clients rather than the mean. Outliers (malicious updates) have less influence.
- Trimmed mean: Remove the top and bottom k% of updates per coordinate, then average the rest.
- Krum: Select the update that's "closest" to the majority by measuring distances to other updates.
- Bulyan: Combines multiple techniques for stronger robustness guarantees.
These methods work by assuming most clients are honest—if fewer than half are malicious, robust aggregation can still produce a good global model.
Key Takeaway: Byzantine-robust aggregation protects federated learning from malicious clients by using statistical techniques (median, trimming, distance-based selection) that reduce the influence of outlier updates, ensuring the global model remains reliable even when some participants cannot be trusted.