DCS

This repo contains a complete, runnable reference implementation of a DCS-style federated learning simulator:

• DDQL (Double DQN) chooses the action (K, λ) each round
  • K = number of selected clients
  • λ = trade-off between trust and latency in the scoring function
• Trust score updated via EMA from a lightweight local “quality” proxy
• Latency prediction via EMA
• Stable projection + kNN anomaly filtering
  • Uses IncrementalPCA when feasible; otherwise falls back to a fixed random projection

This is intentionally a demo-friendly simulator: it models device heterogeneity, dropouts, and a malicious label-flip attack, while keeping runtime reasonable.

---

Quick start (local)

• We recommend running in a fresh virtual environment with Python 3.x to improve compatibility across package versions and reproducibility.
• The repo uses a src/ layout (e.g., src/dcs/), so you must add src to PYTHONPATH. Such as: !PYTHONPATH=$PWD python -u scripts/run_demo.py --config configs/default.yaml

pip install -r requirements.txt

# Run the main demo (DCS vs FedAvg vs ScoreOnly) + bar charts
python scripts/run_demo.py --config configs/default.yaml

# Run hyperparameter sensitivity sweeps (bars generated from runs, not copied from tables)
python scripts/run_sensitivity.py --config configs/default.yaml --out outputs/sensitivity

Outputs (CSVs + PNGs) go to outputs/ by default.

---

Colab usage

In Colab, you can:

1. Upload this repository as a zip OR git clone it.
2. Run:

!pip -q install -r requirements.txt
!python scripts/run_demo.py --config configs/default.yaml

---

How to change datasets

Edit configs/default.yaml:

DATASET: mnist            # mnist | fashion_mnist | cifar10

Notes:

• mnist and fashion_mnist are faster.
• cifar10 is slower due to a larger CNN.

You can also reduce runtime by setting:

MAX_TRAIN_SAMPLES: 20000
MAX_TEST_SAMPLES: 5000
ROUNDS: 10

---

Hyperparameters you will commonly tune

All hyperparameters live in configs/default.yaml.

Federated setup

• NUM_CLIENTS, NUM_EDGES
• ROUNDS, LOCAL_EPOCHS, BATCH_SIZE
• LR, MOMENTUM

Non-IID partition

• DIRICHLET_ALPHA (α)

Client selection action space

• K_MIN, K_MAX, K_STEP
• LAM_GRID (λ values DDQL can choose)

Trust/latency models

• TRUST_ALPHA (EMA factor)
• LAT_EMA (EMA factor)

Anomaly filter

• PCA_RANK
• ANN_NEIGHBORS
• CONTAMINATION

DDQL

• DDQL_GAMMA (discount γ)
• DDQL_EPS_*, DDQL_TAU, DDQL_LR

---

Hyperparameter sensitivity bars

scripts/run_sensitivity.py runs short experiments across grids and produces bar charts like:

• α (Dirichlet non-IID)
• γ (DDQL discount)
• ANN_NEIGHBORS
• λ-grid sweep (fixed scoring mode)

These bars come from measured runs with the same simulator.

---

Repo layout

• src/dcs/ – core simulator and algorithms
• scripts/run_demo.py – DCS vs baselines + bar charts
• scripts/run_sensitivity.py – hyperparameter sweeps + bar charts
• configs/default.yaml – editable config

---