Evidential Trust-Aware Decentralized Federated Learning
A modular framework for Byzantine-resilient decentralized FL with uncertainty-driven peer evaluation and personalized model aggregation for Wearable IoT.
Rangwala, Sinnott, Buyya - University of Melbourne
What is Murmura?
Murmura is a comprehensive framework for Evidential Trust-Aware Decentralized Federated Learning. It leverages Dirichlet-based uncertainty decomposition to evaluate peer trustworthiness and enable Byzantine-resilient model aggregation for wearable IoT applications.
Evidential Deep Learning
Dirichlet-based models for uncertainty quantification
Byzantine Resilient
Robust against malicious nodes using trust-aware filtering
Wearable IoT
Designed for activity recognition on resource-constrained devices
The Key Insight
Epistemic-aleatoric uncertainty decomposition from Dirichlet-based evidential models directly indicates peer reliability:
High Epistemic Uncertainty (Vacuity)
Indicates insufficient learning or evidence - possibly Byzantine behavior. These peers should be filtered.
High Aleatoric Uncertainty (Entropy)
Reflects inherent data ambiguity - the peer is still trustworthy and should be included in aggregation.
This distinction allows intelligent peer filtering that traditional distance-based methods cannot achieve.
Key Contributions
Murmura introduces novel techniques for trust-aware decentralized federated learning.
Evidential Trust-Aware Aggregation
Novel algorithm leveraging Dirichlet-based uncertainty to identify and filter Byzantine peers via cross-evaluation.
Uncertainty-Driven Personalization
Adaptive self-weighting based on local model confidence, balancing knowledge transfer with personalized learning.
BALANCE-Style Threshold Dynamics
Progressive trust threshold tightening as models converge - starting lenient, becoming stricter over training.
Byzantine Attack Resilience
Robust against Gaussian noise and directed deviation attacks with up to 30% compromised nodes.
Flexible Topologies
Support for ring, fully-connected, Erdos-Renyi, and k-regular network topologies.
Config-Driven Experiments
YAML/JSON configuration for reproducible experiments with CLI and Python API support.
Wearable IoT Datasets
Evaluated on three real-world wearable sensor datasets with natural user heterogeneity.
UCI HAR
- Source
- Smartphone sensors
- Nodes
- 10 subjects
- Activities
- 6 classes
- Features
- 561
PAMAP2
- Source
- Body-worn IMUs
- Nodes
- 9 subjects
- Activities
- 12 classes
- Features
- 4000
PPG-DaLiA
- Source
- Wrist-worn PPG/EDA
- Nodes
- 15 subjects
- Activities
- 7 classes
- Features
- 192
Aggregation Algorithms
Compare against state-of-the-art Byzantine-resilient aggregation methods.
Evidential Trust
Uncertainty-aware
FedAvg
Baseline averaging
Krum
Distance-based
BALANCE
Adaptive threshold
Sketchguard
Sketch compression
UBAR
Two-stage robust
Stay Updated
Subscribe to receive updates on Murmura's development and new releases.
We respect your privacy. Unsubscribe at any time.