Deployment

Deploy scenarios using NEBULA

Participant Details

Confirm the deployment of the defined scenario

Information

Scenario Generation

Advanced Deployment

CPU GPU

Aggregation

FedAvg Krum TrimmedMean Median

Topology

Fully Ring Star

Dataset

MNIST FashionMNIST Covtype KDDCUP99 AdultCensus BreastCancer CIFAR10

Robustness

No Attack Label Flipping Sample Poisoning Model Poisoning

GLL Neuron Inversion Swapping Weights Delayer

Discovered VPN Devices

Scenario Information

Scenario title

Scenario description

Deployment

Deployment type

Processes

Virtual devices

Physical devices

Federation Architecture

Federation approach

Number of rounds

Network Topology

Topology generation

Custom topology

Predefined topology

Edge Probability:

LEGEND

Aggregator

Trainer

Server

Malicious

Number of participants: 3

Dataset

Federated dataset

Dataset type

Partition methods

Parameter setting

Training

Model

Aggregation

Aggregation algorithm

Participants

Logging

Reporting

Individual participants

Advanced Topology

Distance between participants

Advanced Training

Number of Epochs

Differential Privacy

Enable/Disable DP

Noise multiplier

Max Grad Norm

Other DP parameters are still taken from nebula/frontend/config/participant.json.example.

Feature Squeezing

Enable/Disable Feature Squeezing

Feature Squeezing is currently available only for image datasets.

n (bit depth)

Applies only to image datasets.

Adversarial Training

Enable/Disable Adversarial Training

Adversarial Training for tabular datasets currently supports AdultCensus, BreastCancer, Covtype, and KDDCUP99 with constrained PGD.

Attack

Epsilon

Apply probability

Tabular candidate selection

Tabular loss window

Tabular margin window

PGD steps

Training mode

Image datasets use FGSM/PGD. AdultCensus, BreastCancer, Covtype, and KDDCUP99 use constrained PGD for tabular adversarial training.

Robustness

Attack Type

Sample Poisoning is currently available only for image datasets.

% Malicious nodes

% Samples poisoned by each malicious node

% Poisoned noise

Noise type

Targeted

Target label

Target changed label

Layer

Delay

Flooding Factor

Target Percentage

Target Selection Interval

Starting round

Stopping round

Attack interval

Defense

Enable/Disable Reputation

Reputation Initialization

Weighting Factor

Reputation Metrics

Model-based metrics

Model Similarity

Fraction Parameters Changed

Communication-based metrics

Number of Messages

Model Arrival Latency

Total weight must not exceed 1.

Mobility

Select the default location for the participants

Custom location

Random location

Latitude and longitude

Mobility configuration

Disable Mobility

Enable Mobility

Network simulation

Type of mobility

Radius of federation (m)

Scheme of mobility

Each participant moves every (rounds)

% mobile participants

Mobility adaptation

Number of additional participants (deployed during federation)

Connection delay (seconds)

Schema of deployment

Situtational Awareness

Enable/Disable SA

Strict Topology

SA Discovery

Candidate Selector

Model Handler

SA Reasoner

Arbitration Policy

Neighbor Policy

Enable/Disable SAR training

Training Policy

Trustworthiness

Enable/Disable Trustworthiness

Robustness pillar

Resilience to attacks notion

Algorithm robustness notion

Client reliability notion

Privacy pillar

Technique notion

Uncertainty notion

Indistinguishability notion

Fairness pillar

Selection fairness notion

Performance fairness notion

Class distribution notion

Outcome fairness notion

Explainability pillar

Interpretability notion

Post hoc methods notion

Accountability pillar

Factsheet completeness notion

Monitoring notion

Architectural soundness pillar

Client management notion

Optimization notion

Federation management notion

Sustainability pillar

Energy source notion

Hardware efficiency notion

Federation complexity notion