Adaptive spectral trust weighted clipped aggregation for Byzantine resilient topology aware federated learning

Federated learning enables collaborative model training without centralized collection of client data, but its reliability can deteriorate under statistical heterogeneity and Byzantine model-update manipulation. This paper proposes Adaptive Spectral Trust-Weighted Clipped Aggregation (ASTWCA), a topology-aware robust aggregation framework that combines a coordinate-wise median reference, similarity-driven temporal trust, trust-dependent adaptive clipping, and graph-derived spectral weighting to control the influence of unreliable client updates. ASTWCA is evaluated on MNIST, FashionMNIST, and CIFAR-10 against FedAvg, Median, Trimmed Mean, Krum, and FLTrust under clean training and sign-flipping, gradient-scaling, Gaussian-noise, and label-flipping attacks. The results show that ASTWCA avoids the severe degradation exhibited by conventional averaging under direct update attacks while remaining competitive with established robust aggregators across datasets. Under sign flipping, ASTWCA achieves 98.42%, 86.83%, and 49.00% accuracy on MNIST, FashionMNIST, and CIFAR-10, respectively, compared with 9.85%, 11.98%, and 10.06% for FedAvg. A dedicated non-IID analysis further shows positive honest–Byzantine trust separation and Byzantine aggregation-weight mass below the nominal 20% adversarial participation fraction across Dirichlet concentrations from 0.1 to 1.0, although severe heterogeneity increases benign-client clipping. Dynamic-topology experiments show negligible accuracy variation across the tested graph-change schedules, while the aggregation-level scalability analysis extends to 500 participating clients. Parameter sensitivity and component ablation identify a robustness–utility trade-off associated with clipping and show that adaptive clipping provides the largest measured contribution against sign flipping. Defense-aware experiments additionally expose a limitation: under the Adaptive-0.25 configuration, the trust gap becomes negative and Byzantine weight mass increases to 0.5491, demonstrating that similarity-based trust can be compromised by reference-compatible adversarial updates. These results characterize ASTWCA as a competitive and interpretable Byzantine-robust aggregation framework while explicitly identifying the conditions under which its trust mechanism can fail.

Authors

Institutions

Publication Details

Journal
Discover Computing
Published
2026-11-17
DOI
https://doi.org/10.1007/s10791-026-10714-9
Primary Topic
Privacy-Preserving Technologies in Data
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Adaptive spectral trust weighted clipped aggregation for Byzantine resilient topology aware federated learning

Suresh Kumar Jha, Niroj Kumar Pani, Aditya Narayan Hati, Ketan Anand
2 links
Discover Computing
Privacy-Preserving Technologies in Data
article

Adaptive spectral trust weighted clipped aggregation for Byzantine resilient topology aware federated learning

Suresh Kumar Jha, Niroj Kumar Pani, Aditya Narayan Hati, Ketan Anand
article en

Abstract

Federated learning enables collaborative model training without centralized collection of client data, but its reliability can deteriorate under statistical heterogeneity and Byzantine model-update manipulation. This paper proposes Adaptive Spectral Trust-Weighted Clipped Aggregation (ASTWCA), a topology-aware robust aggregation framework that combines a coordinate-wise median reference, similarity-driven temporal trust, trust-dependent adaptive clipping, and graph-derived spectral weighting to control the influence of unreliable client updates. ASTWCA is evaluated on MNIST, FashionMNIST, and CIFAR-10 against FedAvg, Median, Trimmed Mean, Krum, and FLTrust under clean training and sign-flipping, gradient-scaling, Gaussian-noise, and label-flipping attacks. The results show that ASTWCA avoids the severe degradation exhibited by conventional averaging under direct update attacks while remaining competitive with established robust aggregators across datasets. Under sign flipping, ASTWCA achieves 98.42%, 86.83%, and 49.00% accuracy on MNIST, FashionMNIST, and CIFAR-10, respectively, compared with 9.85%, 11.98%, and 10.06% for FedAvg. A dedicated non-IID analysis further shows positive honest–Byzantine trust separation and Byzantine aggregation-weight mass below the nominal 20% adversarial participation fraction across Dirichlet concentrations from 0.1 to 1.0, although severe heterogeneity increases benign-client clipping. Dynamic-topology experiments show negligible accuracy variation across the tested graph-change schedules, while the aggregation-level scalability analysis extends to 500 participating clients. Parameter sensitivity and component ablation identify a robustness–utility trade-off associated with clipping and show that adaptive clipping provides the largest measured contribution against sign flipping. Defense-aware experiments additionally expose a limitation: under the Adaptive-0.25 configuration, the trust gap becomes negative and Byzantine weight mass increases to 0.5491, demonstrating that similarity-based trust can be compromised by reference-compatible adversarial updates. These results characterize ASTWCA as a competitive and interpretable Byzantine-robust aggregation framework while explicitly identifying the conditions under which its trust mechanism can fail.

Discover ComputingVol. 29(1)
Indira Gandhi Institute of Technology (IN), Manipal University Jaipur (IN), Sharda University (IN), University of Rajasthan (IN)
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon