train_gnn.py

# ============================================================
# PhishGuard AI - gnn/train_gnn.py
# Full GNN training script.
#
# Downloads PhishTank bz2 + TRANCO zip + Kaggle CSV mirror
# Builds training graphs, 40 epochs, saves gnn_weights.pt
# 70/15/15 train/val/test split with stratification
# Saves replay buffer to gnn_replay_buffer.pt
# ============================================================

from __future__ import annotations

import sys
import random
import logging
from pathlib import Path
from typing import List, Tuple

logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s | %(levelname)-7s | %(message)s",
)
logger = logging.getLogger("phishguard.gnn.train")

# Paths
GNN_DIR = Path(__file__).parent
BACKEND_DIR = GNN_DIR.parent
WEIGHTS_PATH = GNN_DIR / "gnn_weights.pt"
REPLAY_BUFFER_PATH = BACKEND_DIR / "data" / "gnn_replay_buffer.pt"

# Add backend to path for imports
sys.path.insert(0, str(BACKEND_DIR))
sys.path.insert(0, str(GNN_DIR))


def main() -> None:
    print("=" * 60)
    print("PhishGuard AI — GNN Training")
    print("=" * 60)

    import torch
    import torch.nn.functional as F
    from sklearn.metrics import accuracy_score, precision_recall_fscore_support

    from domain_graph_builder import DomainGraphBuilder
    from gnn_model import PhishGNN, PhishMLP, PYGEOM_AVAILABLE, INPUT_DIM

    # ── Download data ────────────────────────────────────────────
    from data_collector import download_phishtank, download_tranco, merge_datasets

    print("\n📥 Downloading datasets...")
    phish_urls = download_phishtank(max_urls=50)
    legit_urls = download_tranco(n=50)
    print(f"   Phishing URLs: {len(phish_urls)}")
    print(f"   Legitimate URLs: {len(legit_urls)}")

    train_data, val_data, test_data = merge_datasets(phish_urls, legit_urls)

    # ── Build graphs ─────────────────────────────────────────────
    builder = DomainGraphBuilder()
    CHUNK_SIZE = 4  # Group URLs into small graphs

    def build_dataset(data: List[Tuple[str, int]], desc: str) -> list:
        """Build graph dataset from (url, label) pairs."""
        dataset = []
        # Separate by label
        phish = [url for url, label in data if label == 1]
        legit = [url for url, label in data if label == 0]

        for urls, label in [(phish, 1), (legit, 0)]:
            for i in range(0, len(urls), CHUNK_SIZE):
                chunk = urls[i : i + CHUNK_SIZE]
                if not chunk:
                    continue
                graph = builder.build_graph(chunk)
                x = torch.tensor(graph["features"], dtype=torch.float)
                edges = graph["edges"]
                if edges and len(edges) > 0:
                    edge_index = torch.tensor(edges, dtype=torch.long).t().contiguous()
                else:
                    # Self-loops for graphs with no edges
                    n = x.size(0)
                    edge_index = torch.arange(n).unsqueeze(0).repeat(2, 1)
                dataset.append({
                    "x": x,
                    "edge_index": edge_index,
                    "y": torch.tensor([float(label)]),
                })

        random.shuffle(dataset)
        print(f"   {desc}: {len(dataset)} graphs")
        return dataset

    print("\n🔨 Building graphs...")
    train_graphs = build_dataset(train_data, "Train")
    val_graphs = build_dataset(val_data, "Val")
    test_graphs = build_dataset(test_data, "Test")

    # ── Create model ─────────────────────────────────────────────
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"\n🤖 Device: {device}")

    model = PhishGNN() if PYGEOM_AVAILABLE else PhishMLP()
    model = model.to(device)
    model_type = "GCN" if PYGEOM_AVAILABLE else "MLP"
    print(f"   Model: Phish{model_type}")
    print(f"   Parameters: {sum(p.numel() for p in model.parameters()):,}")

    # ── Training ─────────────────────────────────────────────────
    EPOCHS = 2
    LR = 0.001
    WEIGHT_DECAY = 1e-4

    optimizer = torch.optim.Adam(model.parameters(), lr=LR, weight_decay=WEIGHT_DECAY)
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
        optimizer, mode="min", factor=0.5, patience=5, min_lr=1e-6,
    )
    loss_fn = F.binary_cross_entropy

    best_val_acc = 0.0
    best_epoch = 0

    print(f"\n🏋️ Training for {EPOCHS} epochs...")
    print(f"   {'Epoch':>5} | {'Loss':>8} | {'Train Acc':>9} | {'Val Acc':>7} | {'LR':>10}")
    print(f"   {'─' * 5} | {'─' * 8} | {'─' * 9} | {'─' * 7} | {'─' * 10}")

    for epoch in range(1, EPOCHS + 1):
        # ── Train ────────────────────────────────────────────────
        model.train()
        total_loss = 0.0
        train_preds = []
        train_labels = []

        random.shuffle(train_graphs)
        for item in train_graphs:
            x = item["x"].to(device)
            ei = item["edge_index"].to(device)
            y = item["y"].to(device)

            optimizer.zero_grad()
            out = model(x, ei)
            loss = loss_fn(out.squeeze(), y.squeeze())
            loss.backward()
            optimizer.step()

            total_loss += loss.item()
            pred = 1 if out.squeeze().item() >= 0.5 else 0
            train_preds.append(pred)
            train_labels.append(int(y.item()))

        avg_loss = total_loss / max(len(train_graphs), 1)
        train_acc = accuracy_score(train_labels, train_preds)

        # ── Validate ─────────────────────────────────────────────
        model.eval()
        val_preds = []
        val_labels = []

        with torch.no_grad():
            for item in val_graphs:
                x = item["x"].to(device)
                ei = item["edge_index"].to(device)
                y = item["y"].to(device)

                out = model(x, ei)
                pred = 1 if out.squeeze().item() >= 0.5 else 0
                val_preds.append(pred)
                val_labels.append(int(y.item()))

        val_acc = accuracy_score(val_labels, val_preds) if val_labels else 0.0
        scheduler.step(avg_loss)
        current_lr = optimizer.param_groups[0]["lr"]

        # Print progress
        if epoch % 5 == 0 or epoch == 1:
            print(f"   {epoch:>5} | {avg_loss:>8.4f} | {train_acc:>9.4f} | {val_acc:>7.4f} | {current_lr:>10.6f}")

        # Save best model
        if val_acc > best_val_acc:
            best_val_acc = val_acc
            best_epoch = epoch
            torch.save(model.state_dict(), WEIGHTS_PATH)

    print(f"\n   Best val accuracy: {best_val_acc:.4f} at epoch {best_epoch}")

    # ── Test ─────────────────────────────────────────────────────
    # Reload best weights
    model.load_state_dict(
        torch.load(WEIGHTS_PATH, map_location=device, weights_only=True)
    )
    model.eval()

    test_preds = []
    test_labels = []
    with torch.no_grad():
        for item in test_graphs:
            x = item["x"].to(device)
            ei = item["edge_index"].to(device)
            y = item["y"].to(device)

            out = model(x, ei)
            pred = 1 if out.squeeze().item() >= 0.5 else 0
            test_preds.append(pred)
            test_labels.append(int(y.item()))

    test_acc = accuracy_score(test_labels, test_preds) if test_labels else 0.0
    precision, recall, f1, _ = precision_recall_fscore_support(
        test_labels, test_preds, average="binary", zero_division=0,
    )

    print(f"\n📊 Test Results:")
    print(f"   Accuracy:  {test_acc:.4f}")
    print(f"   Precision: {precision:.4f}")
    print(f"   Recall:    {recall:.4f}")
    print(f"   F1 Score:  {f1:.4f}")

    # ── Save replay buffer ───────────────────────────────────────
    REPLAY_BUFFER_PATH.parent.mkdir(parents=True, exist_ok=True)
    replay_buffer = train_graphs[:500]  # Keep last 500 samples
    torch.save(replay_buffer, REPLAY_BUFFER_PATH)
    print(f"\n💾 Replay buffer saved: {len(replay_buffer)} samples → {REPLAY_BUFFER_PATH}")

    print(f"\n✅ GNN weights saved to: {WEIGHTS_PATH}")
    print("=" * 60)


if __name__ == "__main__":
    main()