Gossip-based decentralized federated learning (DFL) achieves generalization through iterative pairwise weight averaging, propagating learned representations across all nodes regardless of local data composition. This same mechanism creates a structural failure mode for class unlearning: gossip convergence drives all nodes toward a shared consensus model, so every node's weights encode forget-class knowledge irrespective of local shard composition, yet each must erase it using only local gradient signal. Naive gradient-ascent unlearning fails disproportionately across nodes under non-IID Dirichlet-partitioned data, appearing to succeed by global forget-class accuracy while leaving measurable per-node membership-inference disparity that grows monotonically with heterogeneity. We term this the gossip convergence unlearning paradox and show that per-node MIA AUC variance and per-node L2 distance to a FedRetrain baseline are necessary verification metrics that global means conceal. At extreme concentration (alpha = 0. 1), zero-sample nodes present a structural verification problem: they cannot supply local members for MIA evaluation yet encode the forget-class representation as completely as any other node. We propose Class-Frequency-Aware Unlearning (CFAU), a fully adaptive protocol requiring no inter-node communication beyond standard gossip rounds, and validate all findings on a physical six-node Raspberry Pi Zero 2W cluster executing pure NumPy gradient computation on CIFAR-10 under IID and Dirichlet alpha in 1. 0, 0. 5, 0. 1. Statistical scaling is validated on an extended 30-node cluster, confirming that sigmaMIA scales monotonically with heterogeneity (0. 024 IID to 0. 063 at alpha = 0. 1) and that 21 of 30 nodes are structurally excluded from MIA verification at extreme concentration. CFAU reduces per-node MIA variance by 24-36% under moderate heterogeneity at no additional wall-clock cost relative to naive interleaved.
Paul Taylor (2026) studied this question.