An adaptive federated-learning system called pFedMARL reported substantially lower reconstruction MSE and higher classification F1 than FedAvg in each of three simulated data-partition settings. On the server model, evaluated against all client test sets, pFedMARL’s F1 means were 0.22 in CS, 0.38 in LS and 0.61 in QS, with MSE means of 0.18, 0.24 and 0.24. The centralized baseline scored 0.97 on F1 and 0.01 on MSE.
The finding is narrower than a general performance claim. The authors say pFedMARL may outperform FedAvg, Ditto and local-only training in adversarial scenarios and may match Ditto in non-adversarial settings, with improved fairness and robustness. Those conclusions describe the tested simulation; they do not establish causal superiority or performance beyond it.
An adaptive answer to uneven data
pFedMARL uses a server-side deep-reinforcement-learning agent to adjust the weights given to client updates during aggregation. Client-side agents balance local learning against global knowledge. Both are trained online, without pre-training, using Twin Delayed DDPG. The design targets federated settings in which clients have heterogeneous, non-IID data distributions rather than the same mix of examples.
The test used 10% of the DCASE challenge Task 2 development dataset. It covered 14 machine classes, with recordings sampled at 16 kHz and lasting from 6 to 18 seconds. For each class, the training material included 990 source-domain normal clips and 10 target-domain normal clips; the test material included 200 clips, split into 100 normal and 100 anomalous recordings. The model was a small audio spectrogram transformer trained for patch reconstruction and classification in a semi-supervised federated task.
Researchers divided the data among 15 clients in each of three settings: QS with a partition parameter of 5.0, LS with 0.4 and CS with 2.0. The comparison included FedAvg, Ditto with a tested regularization value of 0.5, local-only training and a centralized oracle. It ran for 100 federated-learning rounds and tracked reconstruction MSE, classification F1, across-client standard deviation and the l2 distance between client and global models.
The gap depends on the comparison
In QS, pFedMARL’s mean MSE was 0.10 on local data and 0.11 on global data. Its F1 means were 0.77 and 0.73, with standard deviations of 0.05 and 0.02. FedAvg recorded MSE means of 1.17 on both evaluations and F1 means of 0.17 on both.
LS showed a similar separation from FedAvg. pFedMARL reported MSE means of 0.10 locally and 0.11 globally, with F1 means of 0.87 and 0.60. FedAvg’s MSE means were 0.96 on both evaluations and its F1 means were 0.13. But local-only training produced an F1 of 0.92 locally and 0.59 globally, so pFedMARL did not lead every comparison on every measure.
The CS setting produced pFedMARL MSE means of 0.06 locally and 0.12 globally, with F1 means of 0.96 and 0.21. FedAvg’s corresponding MSE means were 1.25 and its F1 means were 0.02. Local-only training reached 0.98 locally and 0.21 globally, leaving pFedMARL ahead of FedAvg but slightly below local-only training on the local F1 measure and level with it globally.
The reported pFedMARL F1 standard deviations were 0.05 locally and 0.02 globally in QS, 0.05 and 0.08 in LS, and 0.04 and 0.01 in CS. These are descriptive across-client figures. The study did not report confidence intervals or inferential tests, so the tables show the observed benchmark differences without establishing how they would hold across independent reruns.
A simulated attack, not a real deployment
For the robustness condition, two clients applied additive Gaussian noise with the reported sigma-squared setting of 0.5 to their transmitted updates, while their local models remained unchanged. In the adversarial CS dynamics, benign-client contribution stabilized near 0.5, adversarial influence fell to approximately 0.1, and benign clients reduced their reliance on the server after about 50 rounds.
The trajectories therefore showed down-weighting of adversarial influence, not its complete removal. The authors’ broader interpretation is that this adaptive behavior may improve robustness and fairness in the tested adversarial scenario, but the evidence remains a simulated benchmark comparison rather than a causal estimate of real-world performance.
A useful benchmark with a narrow reach
The evaluation was limited to one 10% subset of one machine-sound dataset and one audio spectrogram-transformer task. Its non-IID partitions and adversarial behavior were simulated, and the federation used 15 clients per scenario. The paper identifies testing on other data modalities and scaling to significantly more clients as future research.
The document is an arXiv preprint, version 1, dated 26 Aug 2026. For now, the evidence is tied to this audio benchmark and its simulated federation, leaving open whether the reported score differences would transfer to larger or different applications.
Paper data and sources
Original title: Cooperative Multi-Agent Reinforcement Learning for Adaptive Aggregation in Semi-Supervised Federated Learning with non-IID Data
Authors: Rene Glitza, Luca Becker, Rainer Martin
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: 10.1109/icassp55912.2026.11462172
Original paper · Full text