app.py

from fastapi import FastAPI
from pydantic import BaseModel
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

app = FastAPI()

# NEW MODEL: Multimodal Phishing Detector (URLs, SMS, Email)
MODEL_ID = "ealvaradob/bert-finetuned-phishing"

print("Loading model... This might take a minute as it's a 'large' BERT model.")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)

class URLInput(BaseModel):
    url: str

@app.get("/")
async def root():
    return {"status": "URL Phishing Detector API is running"}

@app.post("/predict")
async def predict_url(data: URLInput):
    # 1. Basic Pre-check
    if not data.url or len(data.url) < 4:
        return {"error": "Invalid URL provided"}

    # 2. Tokenize and Predict
    inputs = tokenizer(data.url, return_tensors="pt", truncation=True, max_length=512)
    
    with torch.no_grad():
        outputs = model(**inputs)
        # Apply Softmax to get percentages
        predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    
    probs = predictions[0].tolist()
    
    # 3. Dynamic Label Mapping
    # The model usually uses LABEL_0 (Legitimate) and LABEL_1 (Phishing)
    confidences = {model.config.id2label[i]: prob for i, prob in enumerate(probs)}
    
    # Identify the highest confidence label
    max_label = max(confidences.items(), key=lambda x: x[1])
    label_name = max_label[0]
    
    # Check for "LABEL_1" or "phishing" keyword in the output
    is_phishing = "1" in label_name or "phishing" in label_name.lower()

    return {
        "url": data.url,
        "prediction": "phishing" if is_phishing else "legitimate",
        "confidence": round(max_label[1], 4),
        "raw_scores": confidences,
        "is_malicious": is_phishing
    }

if __name__ == "__main__":
    import uvicorn
    # 7860 is the standard port for Hugging Face Spaces
    uvicorn.run(app, host="0.0.0.0", port=7860)