Robust Aggregation in federated learning is the use of Byzantine-resilient aggregation rules to combine client updates — replacing simple averaging (which is vulnerable to a single malicious client) with robust statistics that tolerate a fraction of corrupted or adversarial updates.

Robust Aggregation Methods

• Coordinate-Wise Median: Take the median of each gradient coordinate across clients.
• Trimmed Mean: Remove the highest and lowest values for each coordinate, then average.
• Krum/Multi-Krum: Select the update(s) closest to the majority of other updates.
• Bulyan: Combine Krum selection with trimmed mean for stronger robustness.

Why It Matters

• Byzantine Resilience: Tolerates up to $f < n/2$ malicious or faulty clients (depending on the method).
• Poisoning Defense: Robust aggregation is the primary defense against federated learning poisoning attacks.
• No Accuracy Loss: With few Byzantine clients, robust aggregation matches FedAvg performance.

Robust Aggregation is majority rules, outliers rejected — using robust statistics to aggregate client updates while ignoring adversarial or corrupt contributions.