EthPhishGraph-2026

Ego-graph K=2 dataset of Ethereum phishing accounts for Graph Neural Network based phishing detection and adversarial robustness research.

📊 Dataset Statistics

	Count	% Phishing
Total targets	7,036	50.0%
Phishing	3,518	-
Normal	3,518	-

Temporal split (60/20/20)

Partition	Count	Phishing ratio	Period
Train	4,221	56.8%	≤ T1 = 2018-04-23
Val	1,407	37.3%	T1 → T2 = 2018-07-12
Test	1,408	42.2%	> T2

Graph size

Partition	Avg nodes	Avg edges
Train	2,318	3,241
Val	2,482	3,482
Test	3,677	5,129

hop-2 chiếm ~99% số node mỗi graph.

📁 Dataset Structure

EthPhishGraph-2026/
├── README.md                          # This file
├── LICENSE                            # CC-BY-4.0
├── ego_graph_k2_dataset.pkl           # 4.5 GB — Main pickle (train/val/test splits)
├── temporal_split.csv                 # 449 KB — Address → (label, partition, anchor_ts)
├── temporal_split_meta.json           # — Split metadata (T1, T2, counts)
├── per_target_neighbors.pkl           # 11 MB — Precomputed hop-1 neighbor lists
├── unique_hop1_neighbors.csv          # — Deduplicated neighbor addresses
├── feat_stats.pt                      # — Normalization stats (mean, std) from TRAIN only
├── bad_files.txt                      # — Known bad hop-2 files to skip
├── raw_txs/                           # Raw transaction JSONs (for Stage 6 attacks)
│   ├── phishing_hop1_txs.tar.gz       # ~50 MB — Phishing accounts' hop-1 txs
│   ├── normal_hop1_txs.tar.gz         # ~100 MB — Normal accounts' hop-1 txs
│   └── hop2_neighbor_txs.tar.gz       # ~5 GB — Hop-2 neighbor txs (compressed from 20 GB)
└── misc/                              # Optional artifacts (not needed for basic use)
    ├── stage1_outputs.tar.gz
    ├── stage4_imbalanced.tar.gz
    └── stage8_crosscheck_D2.tar.gz

🔬 Data Schema

ego_graph_k2_dataset.pkl (main file)

import pickle
with open("ego_graph_k2_dataset.pkl", "rb") as f:
    data = pickle.load(f)
# data = {"train": [...], "val": [...], "test": [...]}

Each entry is a dict:

Key	Type	Shape	Description
target	str	—	Ethereum address 0x...
label	int	—	0 = normal, 1 = phishing
target_idx	int	—	Position of target in node list
n_nodes	int	—	Total nodes in the ego-graph
n_edges	int	—	Total directed edges
n_hop1	int	—	Number of hop-1 neighbors
node_features	np.ndarray	[N, 11]	See Node features below
edge_index	np.ndarray	[2, E]	COO format (source, destination)
edge_attr	np.ndarray	[E, 4]	See Edge features below
cutoff	int	—	Unix timestamp used as temporal cutoff

Node features (11 columns)

Col	Name	Description
0	in_count	Number of incoming transactions
1	out_count	Number of outgoing transactions
2	in_value_sum	Total incoming value (wei)
3	out_value_sum	Total outgoing value (wei)
4	in_value_mean	Mean incoming value (wei)
5	out_value_mean	Mean outgoing value (wei)
6	first_ts	First tx timestamp (⚠️ absolute — recommend to drop)
7	last_ts	Last tx timestamp (⚠️ absolute — recommend to drop)
8	active_span	last_ts - first_ts (relative, safe)
9	unique_in_peers	Number of distinct senders
10	unique_out_peers	Number of distinct recipients

Note: Columns 6, 7 are absolute timestamps which may leak temporal patterns to the model (phishing/normal classes have different time distributions). We recommend using only columns [0,1,2,3,4,5,8,9,10] as in NODE_COLS = [0,1,2,3,4,5,8,9,10].

Edge features (4 columns)

Col	Name	Description
0	value	Transaction value (wei)
1	timestamp	Transaction timestamp (⚠️ absolute — recommend to drop)
2	gas_price	Gas price (wei)
3	gas_used	Gas used (units)

Note: We recommend using only EDGE_COLS = [0, 2, 3] to avoid temporal leakage.

📝 Usage Example

import pickle
import numpy as np
import torch
from torch_geometric.data import Data

# Load
with open("ego_graph_k2_dataset.pkl", "rb") as f:
    data = pickle.load(f)
print(f"Train: {len(data['train'])}, Val: {len(data['val'])}, Test: {len(data['test'])}")

# Convert one graph to PyG format
g = data['train'][0]
NODE_COLS = [0, 1, 2, 3, 4, 5, 8, 9, 10]  # drop absolute timestamps
EDGE_COLS = [0, 2, 3]

d = Data(
    x=torch.from_numpy(g['node_features'][:, NODE_COLS]).float(),
    edge_index=torch.from_numpy(g['edge_index']).long(),
    edge_attr=torch.from_numpy(g['edge_attr'][:, EDGE_COLS]).float(),
    y=torch.tensor([g['label']], dtype=torch.long),
)
print(f"Target: {g['target']}  label={g['label']}  nodes={g['n_nodes']}  edges={g['n_edges']}")

🛠️ Pipeline Stages (how the dataset was built)

1. Stage 1 — Collect: Crawl phishing addresses from Etherscan + known phishing lists
2. Stage 2 — Normal crawl: Collect normal (non-phishing) address candidates
3. Stage 3 — Temporal split: Split addresses into train/val/test by timestamp (T1, T2)
4. Stage 9 — K=2 ego-graph construction: For each target, extract hop-1 neighbors and hop-2 neighbors with full history up to the temporal cutoff. Features aggregated per node.
5. Stage 5 — GNN baselines: Train GraphSAGE / GATv2 / PEAE-style on this dataset
6. Stage 6 — Adversarial attacks: Evaluate robustness against 7+ transaction-level attacks
7. Stage 7 — Defenses: Adversarial training + Reliability-aware GNN

📏 Models benchmarked (reference performance)

Macro-F1 on test set (3 seeds mean ± std):

Model	Macro-F1	Phishing Recall	Phishing Precision
GraphSAGE	0.8663 ± 0.0032	0.7772 ± 0.0110	0.9077 ± 0.0054
GATv2	0.8151 ± 0.0108	0.6616 ± 0.0159	0.9068 ± 0.0106
PEAE-style	0.8813 ± 0.0028	0.8283 ± 0.0060	0.8925 ± 0.0072
peaegnnv2	0.8564 ± 0.0040	0.7581 ± 0.0140	0.9035 ± 0.0142

⚔️ Adversarial Robustness Reference

Attack Success Rate (% target flipped) at max 1,200 injected transactions, 3 seeds mean ± std:

Attack	peaegnnv2	GraphSAGE	GATv2
TOH (target → fake wallet)	42.4 ± 3.7	7.1 ± 3.0	18.4 ± 2.5
TIH (fake wallet → target)	16.7 ± 10.3	5.9 ± 1.9	33.2 ± 7.5
BPH-T2P	41.9 ± 5.9	6.2 ± 2.2	14.7 ± 1.9
STC (split + camouflage)	23.1 ± 5.3	8.2 ± 3.0	5.0 ± 1.9
ADAPTIVE-STC-LV	66.6 ± 17.2	15.0 ± 2.0	6.8 ± 1.7

→ Best attack depends on architecture; adaptive-balance mechanism (ADAPTIVE-STC-LV) is strongest on PEAE but weakest on GATv2.

⚠️ Known Limitations

1. Temporal split confound: Phishing/normal class distribution differs across partitions (57% / 37% / 42%) due to their natural temporal distributions. We recommend dropping absolute first_ts, last_ts, timestamp features.
2. Gas price is a time proxy: Gas prices on Ethereum changed over years → models may indirectly use time info even without absolute timestamps.
3. Max 500 txs per neighbor was enforced during Stage 9 to limit RAM.
4. Hop-2 is huge: ~99% of graph nodes are hop-2 → very memory-intensive. Consider sub-sampling for low-RAM systems.
5. Phishing labels derived from publicly reported scam addresses — may contain some noise.

📖 Related Scripts

The code accompanying this dataset is in a separate repository (see src/ folder or thesis repo). Key files:

• stage9_k2_upgrade.py — K=2 ego-graph builder
• stage5_train_baselines.py — GNN training (GraphSAGE, GATv2, PEAE-style)
• stage5_train_baselines_paper.py — Reproduction of Huang et al. 2024 PEAE-GNN
• stage6_adversarial_attacks.py — 7+ transaction-level attacks
• stage7_defense.py — Adversarial training + Reliability-aware GNN

📚 Citation

If you use this dataset, please cite:

@mastersthesis{ethphishgraph2026,
  title = {Evaluating and Enhancing the Robustness of Ethereum Phishing Detection Models
           Against Evasion Attacks on Transaction Graphs},
  author = {[Your Name]},
  school = {[Your University]},
  year = {2026},
  note = {EthPhishGraph-2026 dataset available at https://huggingface.co/datasets/[your-username]/EthPhishGraph-2026}
}

🔗 Related Papers

• Wen et al. 2022, "Hide and Seek: An Adversarial Hiding Approach Against Phishing Detection on Ethereum", IEEE TCSS — source of TOH, TIH, BPH attacks
• Huang et al. 2024, "PEAE-GNN: Phishing Detection on Ethereum via Augmentation Ego-Graph Based on Graph Neural Network", IEEE TCSS — PEAE architecture
• Frankart et al. 2025, "The Feature-Space Illusion", arXiv — related adversarial robustness work on blockchain GNNs

📜 License

CC-BY-4.0 (Creative Commons Attribution 4.0 International).

Ethereum on-chain data is public and permissionless. Phishing/scam labels aggregated from public threat intelligence sources (Etherscan phishing tags, PhishTank, community-reported lists).

🙏 Acknowledgments

Data collected via Etherscan API. Compare dataset D2 from XBlock project (link).