SemLRA: Lightweight Robust Aggregation Against Poisoning Attacks in Semantic Federated Learning

Abstract

Semantic federated learning couples task-oriented semantic communication with distributed training, reducing on-device training burden and end-to-end inference latency. However, poisoning through local data, task labels, or transmitted semantic features can degrade task performance and system reliability. We propose semantic lightweight robust aggregation (SemLRA) for an information-separated dual-server architecture. SemLRA calibrates task losses already generated during split training using the within-round median and median absolute deviation to construct bounded client reliability coefficients. These coefficients jointly reweight client- and server-side model aggregation and regulate subsequent server-side optimization. SemLRA estimates reliability without reference data, client-update inspection, or prior knowledge of the malicious-client count or fraction, while retaining FedAvg-order aggregation complexity. On CIFAR-10, SemLRA maintains accuracy close to its clean-training level under continuous attacks, exceeding the strongest competing defense in mean final-stage accuracy by 3.77 percentage points under label flipping. Under burst attacks, SemLRA provides the strongest accuracy retention across all evaluated scenarios. Measured training runtime shows negligible additional overhead.