Poisoning detection in coded federated learning: impact of linear mixed aggregation on an FLPD-style detector
Year: 2026
Type: article
Abstract: Federated learning (FL) lets clients train a shared model without sharing raw data, but malicious clients can poison it with forged updates whose validity the central server cannot directly verify. Similarity-based detection compares client updates, while coded aggregation reduces uplink cost by mixing updates into packets, whether detection remains effective after such mixing is unclear. This paper evaluates the question on Fashion-MNIST with 100 clients, 10 of them malicious, under two coded aggregation modes and three attacks, namely label flipping, scaled sign reversal and a LIE-style attack. The first two attacks cover mixing sizes m = 1, 2, 5 and 10, the LIE-style variant covers decisive configurations. The LIE-style attack is a purpose-built variant and differs from the published algorithm. In Mode A, decoding precedes detection, which remains lossless in the tested configurations. In Mode B, detection acts on mixed packets and depends on the grouping layout. Under the LIE-style attack at m = 10, individual-level detection coverage is 100%, whereas packet-level coverage is 23% under spread and 100% under consecutive, these units differ, and spread accuracy remains at 86.03%. Under scaled sign reversal at m = 10 with spread, the defense can make the model worse than no defense. Saved traffic can therefore cost detection capability, with outcomes depending on layout as well as mixing size.
Author: Siyuan Yangye
Institution: North China Electric Power University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7234186398
