File size: 24,745 Bytes
6663b5f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
# ============================================================
# PhishGuard AI - main.py
# FastAPI orchestrator β€” Full 4-tier phishing detection pipeline
# with feedback-driven incremental retraining.
#
# Endpoints:
#   POST /analyze       β†’ 4-tier URL phishing analysis
#   POST /analyze/email β†’ BERT-only email body analysis
#   POST /retrain       β†’ Incremental model retraining
#   GET  /model_version β†’ Current model version info
#   GET  /health        β†’ All model load statuses
#
# Architecture:
#   Tier 1: Whitelist O(1) β†’ SAFE exit (~55% traffic)
#   Tier 2: Heuristic 15 signals β†’ BLOCK if >= 80 (~15% blocked)
#   Tier 3: BERT+GNN parallel β†’ BLOCK/SAFE/escalate (~15% exits)
#   Tier 4: CNN visual + brand hash β†’ BLOCK/SAFE (~15% borderline)
# ============================================================

from __future__ import annotations

import os
import sys
import asyncio
import time
import hashlib
import logging
import logging.handlers
from collections import OrderedDict
from contextlib import asynccontextmanager
from pathlib import Path
from typing import List, Optional

from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel

# ── Path setup ────────────────────────────────────────────────────────
BASE_DIR = Path(__file__).parent
for sub_dir in ["gnn", "cnn"]:
    sub_path = BASE_DIR / sub_dir
    if sub_path.is_dir():
        sys.path.insert(0, str(sub_path))

# ── Logging ───────────────────────────────────────────────────────────
log_dir = BASE_DIR / "logs"
log_dir.mkdir(exist_ok=True)

_handler = logging.handlers.RotatingFileHandler(
    log_dir / "phishguard.log",
    maxBytes=5 * 1024 * 1024,
    backupCount=3,
    encoding="utf-8",
)
_handler.setFormatter(logging.Formatter(
    "%(asctime)s | %(levelname)-7s | %(name)s | %(message)s",
    datefmt="%Y-%m-%d %H:%M:%S",
))

logger = logging.getLogger("phishguard")
logger.setLevel(logging.INFO)
logger.addHandler(_handler)
logger.addHandler(logging.StreamHandler())


# ── Import project modules ───────────────────────────────────────────
from url_heuristics import HeuristicScorer, HeuristicResult
from bert_analyzer import BERTPhishingClassifier

# GNN imports
GNN_AVAILABLE = False
gnn_inference = None
try:
    from gnn.gnn_inference import GNNInference
    GNN_AVAILABLE = True
except ImportError:
    try:
        from gnn_inference import GNNInference
        GNN_AVAILABLE = True
    except ImportError:
        logger.warning("GNN module not available")

# CNN imports
CNN_AVAILABLE = False
cnn_inference = None
brand_detector = None
try:
    from cnn.cnn_inference import CNNInference
    from cnn.screenshot_hasher import BrandHashDetector
    from cnn.cnn_model import preprocess_screenshot
    CNN_AVAILABLE = True
except ImportError:
    try:
        from cnn_inference import CNNInference
        from screenshot_hasher import BrandHashDetector
        from cnn_model import preprocess_screenshot
        CNN_AVAILABLE = True
    except ImportError:
        logger.warning("CNN module not available")

from tier3_bert_gnn import Tier3Ensemble
from retraining_service import RetrainingService, FeedbackRecord, RetrainResult


# ── Whitelist (Tier 1) ────────────────────────────────────────────────
WHITELIST: set[str] = {
    "google.com", "youtube.com", "facebook.com", "amazon.com", "wikipedia.org",
    "twitter.com", "instagram.com", "linkedin.com", "microsoft.com", "apple.com",
    "github.com", "stackoverflow.com", "reddit.com", "netflix.com", "paypal.com",
    "bankofamerica.com", "chase.com", "wellsfargo.com", "yahoo.com", "bing.com",
    "outlook.com", "office.com", "live.com", "adobe.com", "dropbox.com",
    "zoom.us", "slack.com", "spotify.com", "twitch.tv", "ebay.com",
    "walmart.com", "target.com", "bestbuy.com", "airbnb.com",
    "x.com", "tiktok.com", "pinterest.com", "quora.com", "medium.com",
}


def get_root_domain(url: str) -> str:
    """Extract root domain from a URL."""
    from urllib.parse import urlparse
    try:
        host = urlparse(url).hostname or ""
        host = host.replace("www.", "")
        parts = host.split(".")
        return ".".join(parts[-2:]) if len(parts) >= 2 else host
    except Exception:
        return ""


# ── URL Cache (LRU, 30-min TTL) ──────────────────────────────────────
CACHE_TTL = 30 * 60
CACHE_MAX = 500


class URLCache:
    def __init__(self, maxsize: int = CACHE_MAX, ttl: int = CACHE_TTL) -> None:
        self._cache: OrderedDict = OrderedDict()
        self._maxsize = maxsize
        self._ttl = ttl

    def get(self, url: str) -> Optional[dict]:
        if url in self._cache:
            entry = self._cache[url]
            if time.time() - entry["ts"] < self._ttl:
                self._cache.move_to_end(url)
                return entry["result"]
            else:
                del self._cache[url]
        return None

    def set(self, url: str, result: dict) -> None:
        self._cache[url] = {"result": result, "ts": time.time()}
        self._cache.move_to_end(url)
        if len(self._cache) > self._maxsize:
            self._cache.popitem(last=False)

    def clear(self) -> None:
        self._cache.clear()


_url_cache = URLCache()


# ── Request/Response Models ───────────────────────────────────────────
class AnalyzeRequest(BaseModel):
    url: str
    heuristic_score: float = 0.0
    page_title: str = ""
    page_snippet: str = ""
    related_urls: list = []


class EmailRequest(BaseModel):
    sender: str
    subject: str = ""
    body: str = ""
    urls: list = []
    timestamp: str = ""


class FeedbackSample(BaseModel):
    url: str
    verdict: str = ""
    confidence: float = 0.0
    tier_used: int = 0
    heuristic_score: int = 0
    signals: list = []
    user_feedback: Optional[str] = None
    timestamp: str = ""
    feedback_ts: Optional[str] = None
    url_hash: str = ""
    session_id: str = ""


class RetrainRequest(BaseModel):
    samples: List[FeedbackSample]
    trigger: str = "count"
    session_id: str = ""
    extension_version: str = ""


# ── Global state ──────────────────────────────────────────────────────
_scorer: Optional[HeuristicScorer] = None
_bert: Optional[BERTPhishingClassifier] = None
_gnn: Optional[GNNInference] = None
_cnn: Optional[CNNInference] = None
_brand: Optional[BrandHashDetector] = None
_tier3: Optional[Tier3Ensemble] = None
_retrain_service: Optional[RetrainingService] = None
_retrain_lock = asyncio.Lock()


# ── Lifespan (startup/shutdown) ───────────────────────────────────────
@asynccontextmanager
async def lifespan(app: FastAPI):
    """Load all models at startup, clean up at shutdown."""
    global _scorer, _bert, _gnn, _cnn, _brand, _tier3, _retrain_service

    logger.info("=== PhishGuard AI starting up ===")

    # Tier 2: Heuristic Scorer
    _scorer = HeuristicScorer()
    logger.info(" Tier 2: HeuristicScorer initialized")

    # Tier 3a: BERT
    _bert = BERTPhishingClassifier()
    _bert.load_model()
    logger.info(" Tier 3a: BERT classifier initialized and loaded")

    # Tier 3b: GNN
    if GNN_AVAILABLE:
        _gnn = GNNInference()
        _gnn.load()
        logger.info(f" Tier 3b: GNN loaded={_gnn.is_loaded}")
    else:
        _gnn = None
        logger.warning(" Tier 3b: GNN not available")

    # Tier 3 Ensemble
    if _gnn:
        _tier3 = Tier3Ensemble(_bert, _gnn)
        logger.info(" Tier 3: Ensemble initialized")
    else:
        _tier3 = None
        logger.warning(" Tier 3: Ensemble not available (GNN missing)")

    # Tier 4: CNN + Brand Detection
    if CNN_AVAILABLE:
        _cnn = CNNInference()
        _cnn.load()
        _brand = BrandHashDetector()
        logger.info(f" Tier 4: CNN loaded={_cnn.is_loaded}, Brand hash DB loaded")
    else:
        _cnn = None
        _brand = None
        logger.warning(" Tier 4: CNN not available")

    # Retraining Service
    _retrain_service = RetrainingService(
        bert_classifier=_bert,
        gnn_inference=_gnn or GNNInference(),
        cnn_inference=_cnn or (CNNInference() if CNN_AVAILABLE else None),
    )
    logger.info(" Retraining service initialized")
    logger.info("=== PhishGuard AI ready ===")

    yield

    logger.info("=== PhishGuard AI shutting down ===")


# ── FastAPI App ───────────────────────────────────────────────────────
app = FastAPI(
    title="PhishGuard AI Backend",
    version="3.0",
    description="4-tier ML phishing detection with feedback-driven retraining",
    lifespan=lifespan,
)

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)


# ── POST /analyze β€” Full 4-tier pipeline ──────────────────────────────
@app.post("/analyze")
async def analyze_endpoint(req: AnalyzeRequest) -> dict:
    """
    Analyze a URL through the 4-tier phishing detection pipeline.

    Tier 1: Whitelist β†’ SAFE
    Tier 2: Heuristic β†’ BLOCK if >= 80
    Tier 3: BERT+GNN ensemble β†’ BLOCK/SAFE/escalate
    Tier 4: CNN visual + brand hash β†’ BLOCK/SAFE
    """
    url = req.url
    details: dict = {}

    # ── TIER 1: Whitelist ────────────────────────────────────────
    root = get_root_domain(url)
    if root in WHITELIST:
        return {
            "url": url,
            "is_phishing": False,
            "confidence": 0.0,
            "method": "whitelist",
            "status": "safe",
            "tier": 1,
            "heuristic_score": 0,
            "signals": [],
            "details": {"whitelisted_domain": root},
        }

    # ── Cache check ──────────────────────────────────────────────
    cached = _url_cache.get(url)
    if cached is not None:
        return cached

    # ── TIER 2: Heuristic scoring ────────────────────────────────
    h_result: HeuristicResult = _scorer.score(url)

    # Use the higher of server-side and browser-side heuristic scores
    h_score = max(h_result.score, int(req.heuristic_score))
    details["heuristic"] = {
        "score": h_result.score,
        "raw_score": h_result.raw_score,
        "signals": h_result.signals,
        "browser_score": int(req.heuristic_score),
        "combined_score": h_score,
    }

    if h_score >= 80:
        result = {
            "url": url,
            "is_phishing": True,
            "confidence": h_score / 100.0,
            "method": "heuristic",
            "status": "blocked",
            "tier": 2,
            "heuristic_score": h_score,
            "signals": h_result.signals,
            "details": details,
        }
        _url_cache.set(url, result)
        logger.info(f"Tier 2 BLOCK | url={url[:60]} | score={h_score}")
        return result

    # ── TIER 3: BERT + GNN Ensemble ──────────────────────────────
    if _tier3 is not None:
        try:
            p3 = await _tier3.predict(
                url=url,
                title=req.page_title,
                snippet=req.page_snippet,
                h_score=h_score,
            )
            details["tier3_score"] = p3
        except Exception as e:
            logger.error(f"Tier 3 error: {e}")
            p3 = h_score / 100.0  # fallback to heuristic
            details["tier3_error"] = str(e)
    else:
        # Tier 3 unavailable β€” use BERT alone + heuristic
        if _bert is not None:
            loop = asyncio.get_event_loop()
            try:
                p_bert = await loop.run_in_executor(
                    None, _bert.predict, url, req.page_title, req.page_snippet,
                )
            except Exception:
                p_bert = 0.5
            h_norm = h_score / 100.0
            p3 = 0.60 * p_bert + 0.40 * h_norm
        else:
            p3 = h_score / 100.0
        details["tier3_score"] = p3
        details["tier3_note"] = "ensemble_unavailable"

    # Tier 3 decision
    decision = Tier3Ensemble.decide(p3)

    if decision == "block":
        result = {
            "url": url,
            "is_phishing": True,
            "confidence": round(p3, 4),
            "method": "bert_gnn_ensemble",
            "status": "blocked",
            "tier": 3,
            "heuristic_score": h_score,
            "signals": h_result.signals,
            "details": details,
        }
        _url_cache.set(url, result)
        logger.info(f"Tier 3 BLOCK | url={url[:60]} | P3={p3:.4f}")
        return result

    if decision == "safe":
        result = {
            "url": url,
            "is_phishing": False,
            "confidence": round(p3, 4),
            "method": "bert_gnn_ensemble",
            "status": "safe",
            "tier": 3,
            "heuristic_score": h_score,
            "signals": h_result.signals,
            "details": details,
        }
        _url_cache.set(url, result)
        logger.info(f"Tier 3 SAFE | url={url[:60]} | P3={p3:.4f}")
        return result

    # ── TIER 4: CNN Visual + Brand Hash (borderline 0.40 ≀ P3 < 0.85)
    if _cnn is not None and _cnn.is_loaded:
        try:
            # Capture screenshot
            screenshot_bytes = await _capture_screenshot_for_tier4(url)

            if screenshot_bytes:
                # CNN prediction
                p_cnn = _cnn.predict(screenshot_bytes)
                details["cnn_prob"] = round(p_cnn, 4)

                # Brand hash check
                brand_boost = 0.0
                if _brand is not None:
                    is_impersonation, brand_name, brand_conf = _brand.detect(
                        screenshot_bytes, url
                    )
                    details["brand"] = {
                        "impersonation_detected": is_impersonation,
                        "brand": brand_name,
                        "confidence": round(brand_conf, 3),
                    }
                    if is_impersonation:
                        brand_boost = 0.25

                # P_final = 0.55Β·P3 + 0.30Β·P_cnn + brand_boost
                p_final = min((0.55 * p3) + (0.30 * p_cnn) + brand_boost, 1.0)
                details["tier4_score"] = round(p_final, 4)

                is_phishing = p_final >= 0.65
                result = {
                    "url": url,
                    "is_phishing": is_phishing,
                    "confidence": round(p_final, 4),
                    "method": "full_ensemble_bert_gnn_cnn",
                    "status": "blocked" if is_phishing else "safe",
                    "tier": 4,
                    "heuristic_score": h_score,
                    "signals": h_result.signals,
                    "details": details,
                }
                _url_cache.set(url, result)
                logger.info(f"Tier 4 {'BLOCK' if is_phishing else 'SAFE'} | url={url[:60]} | P_final={p_final:.4f}")
                return result

        except Exception as e:
            logger.error(f"Tier 4 error: {e}")
            details["tier4_error"] = str(e)

    # Tier 4 unavailable/failed β€” use Tier 3 score with conservative threshold
    is_phishing = p3 >= 0.65
    result = {
        "url": url,
        "is_phishing": is_phishing,
        "confidence": round(p3, 4),
        "method": "bert_gnn_ensemble",
        "status": "blocked" if is_phishing else "safe",
        "tier": 3,
        "heuristic_score": h_score,
        "signals": h_result.signals,
        "details": details,
    }
    _url_cache.set(url, result)
    logger.info(f"Tier 4 fallback β†’ Tier 3 | url={url[:60]} | P3={p3:.4f}")
    return result


async def _capture_screenshot_for_tier4(url: str) -> Optional[bytes]:
    """Capture screenshot for Tier 4 CNN analysis."""
    try:
        from playwright.async_api import async_playwright

        async with async_playwright() as p:
            browser = await p.chromium.launch(headless=True)
            page = await browser.new_page(
                viewport={"width": 1280, "height": 800},
                user_agent=(
                    "Mozilla/5.0 (Windows NT 10.0; Win64; x64) "
                    "AppleWebKit/537.36 Chrome/120.0.0.0 Safari/537.36"
                ),
            )

            # Block heavy resources
            await page.route(
                "**/*.{woff,woff2,ttf,eot,mp4,webm,ogg,wav,mp3}",
                lambda route: route.abort(),
            )

            await page.goto(url, wait_until="domcontentloaded", timeout=10000)
            screenshot = await page.screenshot(type="png")
            await browser.close()
            return screenshot

    except Exception as e:
        logger.warning(f"Tier 4 screenshot failed: {e}")
        return None


# ── POST /analyze/email ───────────────────────────────────────────────
@app.post("/analyze/email")
async def analyze_email_endpoint(req: EmailRequest) -> dict:
    """BERT-only path for email body text analysis."""
    # Sender whitelist check
    sender_domain = req.sender.split("@")[-1].lower() if "@" in req.sender else ""
    if sender_domain in WHITELIST:
        return {
            "status": "safe",
            "analysis": {
                "isPhishing": False,
                "probability": 0.0,
                "reason": "Trusted sender domain",
            },
        }

    # Analyze embedded URLs
    MAX_URLS = 3
    urls_to_check = req.urls[:MAX_URLS]

    if not urls_to_check:
        # Text-only analysis
        if _bert:
            combined = f"{req.subject} {req.body}"
            prob = _bert.predict(combined, req.subject, req.body)
            is_phishing = prob > 0.6
            return {
                "status": "blocked" if is_phishing else "safe",
                "analysis": {
                    "isPhishing": is_phishing,
                    "probability": prob,
                    "reason": "BERT text analysis (no URLs)",
                },
            }
        return {
            "status": "safe",
            "analysis": {
                "isPhishing": False,
                "probability": 0.1,
                "reason": "No URLs and no ML model available",
            },
        }

    # Analyze URLs through the main pipeline
    tasks = [
        analyze_endpoint(AnalyzeRequest(url=u, page_title=req.subject))
        for u in urls_to_check
    ]
    results = await asyncio.gather(*tasks, return_exceptions=True)

    max_prob = 0.0
    phishing_detected = False
    flagged_urls = []

    for idx, r in enumerate(results):
        if isinstance(r, Exception):
            continue
        prob = r.get("confidence", 0.0)
        max_prob = max(max_prob, prob)
        if r.get("is_phishing"):
            phishing_detected = True
            flagged_urls.append(r.get("url", urls_to_check[idx]))

    return {
        "status": "blocked" if phishing_detected else "safe",
        "analysis": {
            "isPhishing": phishing_detected,
            "probability": max_prob,
            "flagged_urls": flagged_urls,
            "reason": "URL analysis via ML ensemble",
        },
    }


# ── POST /retrain β€” Incremental retraining ────────────────────────────
@app.post("/retrain")
async def retrain_endpoint(req: RetrainRequest) -> dict:
    """
    Receive labeled feedback and incrementally update all models.
    Uses asyncio.Lock() to prevent concurrent retraining jobs.
    Timeout: 600s max.
    """
    if _retrain_service is None:
        return {"status": "error", "message": "Retraining service not initialized"}

    # Prevent concurrent retraining
    if _retrain_lock.locked():
        return {
            "status": "skipped",
            "message": "Retraining already in progress",
            "models_updated": [],
        }

    async with _retrain_lock:
        # Convert Pydantic models to FeedbackRecord dataclasses
        records = [
            FeedbackRecord(
                url=s.url,
                verdict=s.verdict,
                confidence=s.confidence,
                tier_used=s.tier_used,
                heuristic_score=s.heuristic_score,
                signals=s.signals,
                user_feedback=s.user_feedback,
                timestamp=s.timestamp,
                feedback_ts=s.feedback_ts,
                url_hash=s.url_hash,
                session_id=s.session_id,
            )
            for s in req.samples
        ]

        try:
            result = await asyncio.wait_for(
                _retrain_service.retrain(records),
                timeout=600,
            )

            # Clear URL cache after retraining (stale results)
            if result.status == "success":
                _url_cache.clear()

            return {
                "status": result.status,
                "models_updated": result.models_updated,
                "samples_used": result.samples_used,
                "duration_seconds": result.duration_seconds,
                "accuracy_delta": result.accuracy_delta,
                "next_retrain_hint": result.next_retrain_hint,
            }

        except asyncio.TimeoutError:
            return {
                "status": "error",
                "message": "Retraining timed out (600s limit)",
            }
        except Exception as e:
            logger.error(f"Retrain endpoint error: {e}")
            return {
                "status": "error",
                "message": str(e),
            }


# ── GET /model_version ────────────────────────────────────────────────
@app.get("/model_version")
async def model_version_endpoint() -> dict:
    """Return current model version info for extension polling."""
    if _retrain_service:
        return _retrain_service.get_version_info()
    return {"version": 0, "updated_at": None, "accuracy": {}}


# ── GET /health ───────────────────────────────────────────────────────
@app.get("/health")
async def health_endpoint() -> dict:
    """Liveness probe with per-tier readiness and model statuses."""
    return {
        "status": "ok",
        "version": "3.0",
        "tier1": True,
        "tier2": _scorer is not None,
        "tier3": _tier3 is not None,
        "tier4": _cnn is not None and _cnn.is_loaded if _cnn else False,
        "retraining_in_progress": _retrain_lock.locked(),
        "model_version": _retrain_service.model_version if _retrain_service else 0,
        "modules": {
            "heuristic": _scorer is not None,
            "bert": _bert is not None and _bert.is_loaded,
            "bert_lazy": _bert is not None and not _bert.is_loaded,
            "gnn": _gnn is not None and _gnn.is_loaded if _gnn else False,
            "cnn": _cnn is not None and _cnn.is_loaded if _cnn else False,
            "brand_hash": _brand is not None,
        },
    }


# ── Legacy feedback endpoint (backward compat) ───────────────────────
@app.post("/feedback")
async def legacy_feedback_endpoint(req: dict) -> dict:
    """Legacy feedback endpoint for backward compatibility."""
    return {"status": "success", "message": "Use POST /retrain for feedback-driven retraining"}


# ── Run directly ──────────────────────────────────────────────────────
# uvicorn main:app --reload --port 8000