A new framework, DW‑FedADMM, adds a performance‑based weight to each device’s local model before it takes part in the global update. The weight is calculated from the device’s most recent accuracy, scaled by the best accuracy seen so far, and is used to blend the current global model with a retained, corrected local state. Because the weight is larger for high‑performing devices, the algorithm favors those models while still enforcing consistency across all devices through a dual‑variable mechanism.

Tests on three malware datasets—Malgenome, TUANDROMD, and Drebin—show that DW‑FedADMM tops the accuracy leaderboard in 9 of 18 tested setups. In the most common scenarios, it achieves a mean accuracy of 95.42 % versus 95.32 % for the baseline FedADMM. At the dataset level, the highest average accuracy on Drebin is 97.64 % ± 0.20 % compared to 97.15 % ± 0.21 % for FedAvg and 97.09 % ± 0.67 % for FedADMM.

Beyond accuracy, DW‑FedADMM produces the lowest overall false‑positive rate, 10.86 %, and the smallest client update norm (0.2097) among all evaluated methods. The latter indicates that individual models change less abruptly, which can help keep communication loads predictable.

All experiments used identically distributed data shards, so the impact of heterogeneous client data remains to be tested. Additionally, the dynamic weight is fixed per round; future work could explore adaptive, event‑triggered updates to further reduce communication while keeping speed.