JabaleNurAdnan/dml-fl-iot-ids

DML-FL IoT IDS Benchmark Results

Results from the paper: Dynamic Multilevel Clustered Federated Learning for IoT Intrusion Detection

Grid (72 total runs)

Dimension	Options
Models	CNN, BiLSTM, CNN+BiLSTM, Transformer (fixed)
FL Algorithms	FedAvg, FedProx, FedNova
Datasets	CICIDS2017, UNSW-NB15, Edge-IIoTset
Distributions	IID, Non-IID (Dirichlet α=0.5)
Clients / Rounds	30 clients / 30 rounds / 3 local epochs

Files

File	Description
results/all_results_merged_72runs.csv	Main table — full 72-run benchmark
results/all_results_transformer_18runs.csv	Fixed Transformer (18 runs)
results/all_results_original_54runs.csv	CNN / BiLSTM / CNN+BiLSTM (54 runs)
history/<tag>__round_metrics.csv	Per-round F1, Acc, AUC, threshold

Key Findings

• CNN+BiLSTM dominates across all datasets (F1: 94–99.99%)
• FedNova edges out FedAvg/FedProx on CICIDS2017 non-IID
• IID vs non-IID gap is minimal for CNN+BiLSTM
• CNN near-matches CNN+BiLSTM at 11× fewer params — efficiency winner
• Transformer fix: added Dense(d_model=32) token-projection before MultiHeadAttention

Transformer Bug & Fix

# BROKEN (615 params, d_model=1, collapses)
attn = MultiHeadAttention(...)(inp, inp)      # inp: (n_features, 1)

# FIXED (~55k params, d_model=32, learns)
x    = Dense(32)(inp)                         # (n_features, 32)
attn = MultiHeadAttention(...)(x, x)