Abstract
<title>Abstract</title> <p>Federated learning (FL) trains a shared model across clients without sharing raw data, but privacy regulations increasingly require unlearning a departing client's influence. Federated unlearning (FU) research optimizes how completely and cheaply a client is erased, but ignores collateral effects on remaining clients: whether removal degrades the model unevenly or amplifies bias against under-represented groups under non-IID data. We study these side-effects in the departed-client setting, where the leaving client no longer participates in its own removal. Using from-scratch retraining as reference, we benchmark FedEraser, projected gradient ascent (PGA), fine-tuning, and continue-to-train on two tabular datasets with sensitive attributes and two image datasets, sweeping Dirichlet heterogeneity, four departed-client profiles, and FedAvg/FedProx, while assessing stability, per-client and group fairness, and verified forgetting. Unlearning is cheap on average but uneven: global accuracy is nearly preserved, while per-client accuracy dispersion and demographic-parity and equalized-odds gaps shift substantially; cost grows with heterogeneity and with the departed client's size and rarity. A backdoor forgetting probe shows that the cheapest methods are cheap because they fail to unlearn, whereas FedEraser and PGA truly remove influence. These findings expose a stability–fairness–forgetting trade-off hidden by aggregate metrics and provide an open evidence base for fairness-aware FU.</p>