Skip to main navigation Skip to search Skip to main content

Hierarchical Validity-Weighted Mitigation of Data Poisoning Attacks in Federated Learning

  • The British University in Egypt

Research output: Contribution to conference › Paper › peer-review

4 Downloads (Pure)

Abstract

Data poisoning defenses in federated learning typically operate at a single level, leaving a gap: a client whose poisoned samples are filtered still contributes its remaining data at full aggregation weight. This paper presents a hierarchical
mitigation framework operating at edge and aggregation levels. Each edge node uses an auxiliary detector (LSTM or SNN) to assign continuous validity weights at sample and dataset levels, with the two tiers composing sequentially to suppress malicious contributions. Across 48 attack configurations on MNIST spanning continuous and periodic label-flipping regimes, LSTM-based variants stay within 1.1 percentage point of the clean baseline, while SNN-based variants exceed baseline by +3.3 to +6.5 pp through a data-cleaning effect during clean rounds. Hierarchical mitigation outperforms its local-only counterpart by +2.9 to +4.8 pp and outperforms Confident Federated Learning in every tested configuration.

Workshop

WorkshopIEEE International Workshop on Computer Aided Modeling and Design of Communication Links and Networks, 2026
Country/TerritoryCyprus
CityNicosia
Period7/09/26 → 9/09/26
Internet address

Cite this