Detection methodology

apps/ml/sybilshield/clustering/funding_cluster.py

• · Group addresses by their first incoming (non-CEX) funder
• · Filter out groups with <3 members
• · Confidence scales by time spread of the funding events: <24h → 0.95, <7d → 0.80, else 0.60
• · CEX hot wallets (Binance, Coinbase, etc.) are explicitly excluded

apps/ml/sybilshield/clustering/behavior_cluster.py

• · Build 10-feature vector per address (timing, entropy, protocol diversity, etc.)
• · Drop zero-variance columns (sigma < 1e-12) to prevent uniform-input failure
• · z-score normalize, then HDBSCAN with min_cluster_size=5, leaf cluster selection
• · Fallback: if HDBSCAN returns all-noise AND ≥3 features have zero variance across the batch, treat the whole batch as one Sybil cohort (real-world catch for uniform-script farms)
• · Intra-cluster spread → confidence (lower spread = higher confidence)

apps/ml/sybilshield/clustering/graph_community.py

• · Build directed weighted graph from in-batch transactions (igraph backend, scales to 500K+ nodes)
• · Run Leiden algorithm via leidenalg (RBConfigurationVertexPartition, resolution=1.0, seed=42)
• · Filter communities by size (5 ≤ n ≤ 500) and density (≥ 0.3)
• · Both size and density bounds matter: dense rings are Sybil, but legitimate DEX-using groups also form communities — density threshold cuts them

apps/ml/sybilshield/clustering/cross_chain.py

• · Deterministic: bridge tx from address A on chain X to address B on chain Y → same entity by construction (Stargate, Hop, Across, Wormhole)
• · Probabilistic: same funder + same timestamp window (±10min) across chains → likely-same entity
• · Union-find aggregates into entity components
• · Cluster fires when ≥3 chain-nodes belong to one entity

apps/ml/sybilshield/features/temporal.py

• · Hour-of-day entropy (humans: > 2.5, bots: < 1.5)
• · Day-of-week entropy
• · Min inter-tx seconds (humans rarely < 60s)
• · Burst score (proportion of txs concentrated in 10% of activity window)
• · Lag-1 autocorrelation of inter-tx gaps (mechanical regularity)
• · Activity ratio (active days / account age days)

apps/ml/sybilshield/scoring/train.py

• · LightGBM binary classifier with class_weight='balanced'
• · Sample-weighted by confidence tier (T1=4.0, T4=0.8, G1=4.0, G2=2.0)
• · Trained on ALL tiers (T1-T5 + G1-G2 with weights)
• · Evaluated on T1+T2 sybil + G1 genuine holdout ONLY - never against T4 detector outputs (which are themselves noisy)
• · Honest metrics published: P@70, R@70, FPR-on-G1, ROC-AUC, adversarial recall

Reproducibility

Every model artifact stores its feature_schema_hash and training_manifest_hash. If you can't reproduce a score from the published artifact + manifest hashes, the score is invalid.