Dual asymmetric momentum improves federated class unlearning in edge systems (Code and Reproducibility Scripts)

This Zenodo record provides the reference implementation of FedDAM, a parameter-efficient post-hoc federated class unlearning method designed for resource-constrained edge deployments. The release includes training and unlearning pipelines for CIFAR-10 and CIFAR-100 under federated non-IID Dirichlet partitions, matched-budget comparisons, and evaluation scripts to regenerate the key tables/figures reported in the associated Scientific Reports submission. Key features:(i) Auxiliary-head-only unlearning with frozen backbone and main classifier to reduce communication and client compute.(ii) Dual-asymmetric momentum buffers to decouple retain vs forget optimization dynamics.(iii) Reproducibility controls: fixed seeds, paired partitions and participation schedules, and configuration files for sweeps. New analyses included for revision:(A) Gradient cosine similarity analysis comparing FedAU vs FedDAM (retain vs forget gradient conflict and update alignment).(C) Forget-set heterogeneity statistics: distribution of ∣Du,k∣|D_{u,k}|∣Du,k∣ and fraction of clients with ∣Du,k∣=0|D_{u,k}|=0∣Du,k∣=0 across αD\alpha_DαD. Dataset: CIFAR-10 and CIFAR-100 (public). No new datasets were generated.How to reproduce: See README.md in the archive.