How saranrajsnkr/swb-dm-full-empirical-study works
SWB-DM is a research implementation and empirical evaluation suite for Byzantine-robust federated learning under partial client participation. In each round, an experiment samples clients; they train locally and may submit poisoned updates. The server either aggregates sampled updates directly or updates a full-pool delayed-momentum cache and aggregates that pool. Experiments evaluate the resulting global model and compare robustness, generalization, and scale. The graph preserves the documented study branches; source excerpts are partial, so unshown runtime links are omitted rather than inferred.
Main entry points: grid_main.py Main experiment grid · aggregators.py FedAvg, Krum, Bulyan, SWB · fltrust_corrected.py FLTrust aggregation · cache.py Full-pool momentum cache
Components 20
Federated training
- Client data partitions
common.py - Local training and attacks
common.py - Adaptive attack generation
attacks.py - Model evaluation
common.py
Robust aggregation
- FedAvg, Krum, Bulyan, SWB
aggregators.py - FLTrust aggregation
fltrust_corrected.py - Full-pool momentum cache
cache.py
Evaluation studies
- Main experiment grid
grid_main.py - Attack validation
experiment_attacks_v2.py - Extended method comparison
extended_all_methods.py - Extended-round comparison
grid_extended_rounds.py - Dataset generalization
generalization.py - 500-client scalability
scalability.py - Krum scale check
Krum_500-client_scale.py - Adaptive attack experiments
experiment_adaptive.py - Attack experiments
experiment_attacks.py - Generalization checks
swb_dm_cifar100_extended.py
Analysis
- Robustness bound analysis
theory_bound.py - Adaptive attack bounds
theory_bound_adaptive_attacks.py
Other components
- Researcher
Connections 30
- Researcher → Main experiment grid: runs study
README.md - Main experiment grid → Client data partitions: loads partitions
grid_main.py - Main experiment grid → Local training and attacks: trains clients
grid_main.py - Main experiment grid → Adaptive attack generation: generates attacks
grid_main.py - Main experiment grid → FedAvg, Krum, Bulyan, SWB: aggregates updates
grid_main.py - Main experiment grid → FLTrust aggregation: calls FLTrust
grid_main.py - Main experiment grid → Full-pool momentum cache: updates cache
grid_main.py - Main experiment grid → Model evaluation: evaluates model
grid_main.py - Local training and attacks → Client data partitions: trains on
common.py - Attack validation → Local training and attacks: trains clients
experiment_attacks_v2.py - Attack validation → FedAvg, Krum, Bulyan, SWB: selects method
experiment_attacks_v2.py - Attack validation → FLTrust aggregation: calls FLTrust
experiment_attacks_v2.py - Attack validation → Full-pool momentum cache: steps cache
experiment_attacks_v2.py - Extended method comparison → Local training and attacks: trains clients
extended_all_methods.py - Extended method comparison → FedAvg, Krum, Bulyan, SWB: compares methods
extended_all_methods.py - Extended method comparison → FLTrust aggregation: calls FLTrust
extended_all_methods.py - Extended method comparison → Full-pool momentum cache: uses cache
extended_all_methods.py - Extended-round comparison → FedAvg, Krum, Bulyan, SWB: compares methods
grid_extended_rounds.py - Extended-round comparison → Full-pool momentum cache: steps cache
grid_extended_rounds.py - Dataset generalization → FedAvg, Krum, Bulyan, SWB: compares methods
generalization.py - Dataset generalization → FLTrust aggregation: calls FLTrust
generalization.py - Dataset generalization → Full-pool momentum cache: steps cache
generalization.py - 500-client scalability → FedAvg, Krum, Bulyan, SWB: compares methods
scalability.py - 500-client scalability → FLTrust aggregation: calls FLTrust
scalability.py - Krum scale check → FedAvg, Krum, Bulyan, SWB: calls Krum
Krum_500-client_scale.py - Adaptive attack experiments → Adaptive attack generation: uses attacks
experiment_adaptive.py - Attack experiments → Adaptive attack generation: uses attacks
experiment_attacks.py - Generalization checks → FedAvg, Krum, Bulyan, SWB: compares methods
swb_dm_cifar100_extended.py - Robustness bound analysis → FedAvg, Krum, Bulyan, SWB: analyzes SWB
theory_bound.py - Adaptive attack bounds → Adaptive attack generation: analyzes attacks
theory_bound_adaptive_attacks.py