MLFCIL: A Multi-Level Forgetting Mitigation Framework for Federated Class-Incremental Learning in LEO Satellites

Heng Zhang; Xiaohong Deng; Sijing Duan; Wu Ouyang; KM Mahfujul; Yiqin Deng; Zhigang Chen

cs.NIMar 14, 2026

MLFCIL: A Multi-Level Forgetting Mitigation Framework for Federated Class-Incremental Learning in LEO Satellites

Authors: Heng Zhang, Xiaohong Deng, Sijing Duan, Wu Ouyang, KM Mahfujul, Yiqin Deng, Zhigang Chen

Organizations: School of Information Engineering, Gannan University of Science and Technology, Ganzhou 341000, China · Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China · Department of Electrical and Computer Engineering, Queen University, Kingston, ON K7L 3N6, Canada · School of Data Science, Lingnan University, Tuen Mun, Hong Kong, China · School of Computer Science and Engineering, Central South University, Changsha 410083, China

Abstract

Low-Earth-orbit (LEO) satellite constellations are increasingly performing on-board computing. However, the continuous emergence of new classes under strict memory and communication constraints poses major challenges for collaborative training. Federated class-incremental learning (FCIL) enables distributed incremental learning without sharing raw data, but faces three LEO-specific challenges: non-independent and identically distributed data heterogeneity caused by orbital dynamics, amplified catastrophic forgetting during aggregation, and the need to balance stability and plasticity under limited resources. To tackle these challenges, we propose MLFCIL, a multi-level forgetting mitigation framework that decomposes catastrophic forgetting into three sources and addresses them at different levels: class-reweighted loss to reduce local bias, knowledge distillation with feature replay and prototype-guided drift compensation to preserve cross-task knowledge, and class-aware aggregation to mitigate forgetting during federation. In addition, we design a dual-granularity coordination strategy that combines round-level adaptive loss balancing with step-level gradient projection to further enhance the stability-plasticity trade-off. Experiments on the NWPU-RESISC45 dataset show that MLFCIL significantly outperforms baselines in both accuracy and forgetting mitigation, while introducing minimal resource overhead.

Figures & tables

Fig. 1: FCIL scenario in LEO: edge satellites perform on-board class-incremental training and synchronize with a regional aggregation satellite via inter-satellite links (ISL), without reliance on ground stations.

Notation	Description
$N$ , $T$ , $C$	Number of clients, tasks, and classes.
$\mathcal{N}$ , $\mathcal{Y}_{t}$ , $\mathcal{Y}_{1:t}$	Client index set; class set of task $t$ ; cumulative class set.
$\mathcal{D}_{i}^{(t)}$	Local dataset at client $i$ for task $t$ .
$n_{i}^{(c)}$ , $N_{c}$	Class- $c$ samples at client $i$ ; total $N_{c}=\sum_{i}n_{i}^{(c)}$ .
$\theta{=}\{\theta_{f},\theta_{c}\}$	Model: feature extractor $\theta_{f}$ and classifier $\theta_{c}$ .
$\theta_{i}$ , $\theta_{G}$ , $\theta_{\text{old}}$	Local model at client $i$ ; global model; teacher model from previous task.

TABLE I: Key Mathematical Notations

Fig. 4: Prototype-guided feature drift compensation. (a) Feature drift problem. (b) After drift compensation.

Fig. 5: Stability-plasticity gradient projection. (a) Before projection (conflict). (b) After projection (resolved).

Mechanism	$\theta_{f}$ ( Extractor )	$\theta_{c}$ ( Classifier )
$\mathcal{L}_{\text{cls}}$ (plasticity)	$\checkmark$ (projected)	$\checkmark$
$\mathcal{L}_{\text{distill}}$ (stability)	$\checkmark$	$\checkmark$
$\mathcal{L}_{\text{replay}}$ (stability)	—	$\checkmark$
Adaptive balancing	adjusts $\lambda$ globally
Gradient projection	$\checkmark$	—
Class-aware aggr.	—	$\checkmark$

TABLE II: Which mechanisms update which parameter group (

\theta_{f}

vs.

\theta_{c}

), and where gradient conflicts are resolved (projection on

\theta_{f}

Method	$A_{1}$	$A_{2}$	$A_{3}$	$\bar{A}$	PD $\downarrow$
FedAvg	78.2	52.4	41.3	57.3	36.9
FedAvg+KD	78.2	60.1	53.6	64.0	24.6
FedAvg+Replay	78.2	61.8	55.2	65.1	23.0
MLFCIL (Ours)	78.2	70.5	67.8	72.2	10.4
Joint Training (Oracle)	78.2	76.8	75.1	76.7	3.1

TABLE III: Cumulative accuracy (%) across incremental tasks on NWPU-RESISC45.

\bar{A}

: average incremental accuracy; PD: performance degradation (

A_{1}{-}A_{T}

). Best results (excluding oracle) are in bold .

Fig. 6: Cumulative accuracy across three incremental tasks for all compared methods.

Configuration	$A_{2}$	$A_{3}$	$\bar{A}$	PD $\downarrow$
FedAvg (none)	52.4	41.3	57.3	36.9
+CW	55.1	44.6	59.3	33.6
+CW+KD	61.8	54.3	64.8	23.9
+CW+KD+MR	64.5	58.7	67.1	19.5
+CW+KD+MR+CA	66.2	61.4	68.6	16.8
+CW+KD+MR+CA+AB	67.8	63.5	69.8	14.7

TABLE IV: Ablation study on the NWPU-RESISC45 dataset. CW: class-reweighted loss; KD: knowledge distillation; MR: memory replay; CA: class-aware aggregation; AB: adaptive balancing; DC: drift compensation; GP: gradient projection. “Full” denotes MLFCIL with all components.

Fig. 7: Ablation study: final accuracy

A_{3}

for each incremental configuration in Table IV .

Fig. 8: Forgetting-aware adaptive balancing dynamics. (a) Forgetting score

\mathcal{F}_{i}^{(r)}

across rounds. (b) Adapted loss weights

\lambda_{\text{distill}}^{(r)}

and

\lambda_{\text{replay}}^{(r)}

Fig. 9: Gradient conflict analysis. (a) Frequency of gradient conflicts (

\cos(\mathbf{g}_{\text{plas}},\mathbf{g}_{\text{stab}})<0

) across training steps. (b) Cosine similarity distribution between plasticity and stability gradients with and without projection.

Fig. 10: Old-class vs. new-class accuracy across tasks for MLFCIL and FedAvg.

Fig. 11: Accuracy comparison between old and new classes across tasks.

Method	Dirichlet $\alpha_{\mathrm{Dir}}$
	0.1	0.3	0.5	1.0	5.0	$\Delta\!\downarrow$
FedAvg	31.7	37.2	41.3	48.6	54.1	22.4
FedAvg+KD	43.8	49.1	53.6	58.3	62.5	18.7
FedAvg+Replay	45.2	50.8	55.2	59.7	63.4	18.2
Ours	58.4	63.7	67.8	71.2	73.5	15.1

TABLE V: Final accuracy

A_{3}

(%) under different Non-IID levels (Dirichlet

\alpha_{\mathrm{Dir}}

\Delta

: accuracy drop from

\alpha_{\mathrm{Dir}}{=}5.0

\alpha_{\mathrm{Dir}}{=}0.1

Fig. 12: Per-class accuracy for six representative classes (two from each of the three tasks), evaluated after all tasks, under Non-IID distributions (

\alpha_{\mathrm{Dir}}{=}0.5

Fig. 13: Impact of memory buffer size on final accuracy

A_{3}

Component	Type	Cost
Communication per round
Model params	bidir.	${\sim}$ 83 MB
Prototypes (45 $\times$ 256)	up+down	+45 KB (+0.1%)
Class counts (45 int.)	upload	+0.2 KB
Computation per step
Teacher forward (KD)	1 $\times$ fwd	+2.5%

TABLE VI: Resource overhead analysis. All overhead percentages are relative to FedAvg.

Explore similar work

FedOrbit: Adaptive Personalized Federated Learning for Non-IID LEO Satellite Constellations
Aug 10, 2026Satwat Bashir, Tasos Dagiuklas, Muddesar IqbalHierarchical Federated Learning Federated Learning Aggregation
School of Engineering London South Bank University, London, UK
Resource-Aware Federated Mixture-of-Experts with Adaptive Pruning for Onboard Learning in LEO Satellite Constellations
Sep 25, 2026Mohamed Shaaban, Mohamed Elmahallawy, Marius Bernahrndt +1Federated Learning Mixture of Experts
School of Engineering and Applied Science, Washington State University, Richland, WA 99354, USA · Institute for Software Technology, German Aerospace Center (DLR), Porz, 51147 Cologne, Germany
ORBITALIF: An Efficient Spiking Federated Learning Framework for Onboard Cloud Removal
Aug 25, 2026Bohan Zhang, Chenyu Xu, Yijie Mao +1Remote Sensing Federated Learning
School of Information Science and Technology, ShanghaiTech University, Shanghai, China

Aug 10, 2026J/K move · Enter open · S save

FedOrbit: Adaptive Personalized Federated Learning for Non-IID LEO Satellite Constellations

Satwat Bashir, Tasos Dagiuklas, Muddesar Iqbal

School of Engineering London South Bank University, London, UK

Hierarchical Federated Learning Federated Learning Aggregation Non-IID Federated Learning LEO Satellite Networks Federated Learning

Open paper AskPDF arXiv

Federated learning (FL) in Low Earth Orbit (LEO) satellite constellations is affected by non-IID data and irregular ground-station visibility, both driven by orbital geometry. Global aggregation performs poorly when orbit-level class distributions are disjoint, while strong personalisation can be excessive when these distributions overlap. We present FedOrbit, which combines continuous orbit-level training over inter-satellite links, class-aware hierarchical aggregation, quality-weighted feature aggregation with return-rate dampening, and adaptive feature decomposition based on inter-orbit class similarity. Across three remote-sensing benchmarks and two non-IID partitions, FedOrbit achieves the highest accuracy in five of six settings and is within

0.9

percentage points of the best result in the sixth. The gains over the strongest baseline reach

16.1

percentage points under Dirichlet partitioning and

8.6

under pathological partitioning, with the smallest per-orbit accuracy spread in five of six settings.

Abstract

Explore similar work

FedOrbit: Adaptive Personalized Federated Learning for Non-IID LEO Satellite Constellations

Resource-Aware Federated Mixture-of-Experts with Adaptive Pruning for Onboard Learning in LEO Satellite Constellations

ORBITALIF: An Efficient Spiking Federated Learning Framework for Onboard Cloud Removal