from fastapi import FastAPI, HTTPException, Security from fastapi.security import APIKeyHeader from pydantic import BaseModel from transformers import BertTokenizerFast, BertForSequenceClassification import torch import torch.nn.functional as F import re import os from collections import Counter app = FastAPI() from fastapi.middleware.cors import CORSMiddleware app.add_middleware( CORSMiddleware, allow_origins=["*"], allow_credentials=True, allow_methods=["*"], allow_headers=["*"], ) API_KEY_NAME = "X-API-Key" api_key_header = APIKeyHeader(name=API_KEY_NAME, auto_error=False) async def verify_api_key(api_key: str = Security(api_key_header)): expected = os.environ.get("API_SECRET_KEY") if not api_key or api_key != expected: raise HTTPException(status_code=403, detail="Forbidden") return api_key MODEL_NAME = "Aitouaouch/bert_model_for_spam_detection" tokenizer = BertTokenizerFast.from_pretrained(MODEL_NAME, token=os.environ.get("HF_TOKEN")) model = BertForSequenceClassification.from_pretrained(MODEL_NAME, token=os.environ.get("HF_TOKEN")) model.eval() STOPWORDS = { "the","a","an","is","it","in","on","at","to","for","of","and","or", "but","be","are","was","were","this","that","with","from","have","has", "you","your","we","our","i","my","he","she","they","their","its","will", "can","not","no","so","do","if","as","by","up","all","more","about","just" } class EmailRequest(BaseModel): subject: str body: str domain: str @app.get("/") def health(): return {"status": "BERT Spam Detection Running ✅"} def extract_keywords(text: str, top_n: int = 8) -> list[str]: words = re.findall(r'\b[a-zA-Z]{4,}\b', text.lower()) filtered = [w for w in words if w not in STOPWORDS] freq = Counter(filtered) return [word for word, _ in freq.most_common(top_n)] @app.post("/predict") async def predict(request: EmailRequest, api_key: str = Security(verify_api_key)): text = f"[DOMAIN] {request.domain} [BODY] {request.subject} {request.body}" inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=256) with torch.no_grad(): outputs = model(**inputs) probs = F.softmax(outputs.logits, dim=1) prediction_idx = torch.argmax(probs, dim=1).item() confidence = probs[0][prediction_idx].item() * 100 prediction = "spam" if prediction_idx == 1 else "safe" keywords = extract_keywords(f"{request.subject} {request.body}") return { "prediction": prediction, "confidence": round(confidence, 2), "keywords": keywords }