codebert_Malvec / app.py
111306047
upload som to hf space
8a33636
Raw History Blame
16.4 kB
import os
CACHE_DIR = "/tmp/hf_cache"
os.environ["HF_HOME"] = CACHE_DIR
os.environ["TRANSFORMERS_CACHE"] = CACHE_DIR
os.environ["HUGGINGFACE_HUB_CACHE"] = CACHE_DIR
os.makedirs(CACHE_DIR, exist_ok=True)
import torch
from fastapi import FastAPI, UploadFile, File, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from typing import List, Optional, Dict
from collections import Counter
import json
from huggingface_hub import hf_hub_download
from som_analyzer import SOMAnalyzer
from datasets import load_dataset
import numpy as np
app = FastAPI()
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_methods=["*"],
allow_headers=["*"],
)
# ===== 模型設定 =====
MODEL_REPO = "raxhel/Malvec_predict"
HF_TOKEN = os.environ.get("hf_token")
DEVICE = torch.device("cpu")
# ===== 載入模型 =====
print("🤗 Loading model...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_REPO, token=HF_TOKEN)
model = AutoModelForSequenceClassification.from_pretrained(
MODEL_REPO,
token=HF_TOKEN,
output_attentions=True, # ✅ 輸出注意力權重
output_hidden_states=True # ✅ 關鍵: 輸出 hidden states
)
model.to(DEVICE)
model.eval()
print("✅ Model loaded successfully")
# ===== 新增:載入 SOM 模型 =====
print("🧠 Initializing SOM Analyzer...")
SOM_REPO = "malvec/codebert_Malvec" #
SOM_AVAILABLE = False
try:
#
som_weights_path = hf_hub_download(
repo_id=SOM_REPO,
filename="som_weights.npy",
token=HF_TOKEN #
)
background_data_path = hf_hub_download(
repo_id=SOM_REPO,
filename="som_background_data.pkl",
token=HF_TOKEN #
)
som_analyzer = SOMAnalyzer(
som_weights_path=som_weights_path,
background_data_path=background_data_path
)
print("✅ SOM Analyzer initialized successfully!")
SOM_AVAILABLE = True
except Exception as e:
print(f"⚠️ SOM Analyzer initialization failed: {e}")
print(" SOM analysis will be disabled")
# ===== 新增:從 main.py 複製過來的輔助函數 =====
# (注意:請從你的 main.py 複製最新的程式碼)
async def get_embeddings_from_hf_dataset(filename: str) -> Optional[List[np.ndarray]]:
# ... (請從 main.py 完整複製貼上) ...
try:
print(f"\n📦 Fetching embeddings from HF dataset for: {filename}")
dataset = load_dataset("lyi029/malvec-embeddings", split="train", streaming=True)
embeddings = []
count = 0
for sample in dataset:
sample_filename = sample.get('source', '') or sample.get('filename', '')
if filename.lower() in sample_filename.lower():
emb = sample.get('embedding') or sample.get('emb')
if emb is not None:
embeddings.append(np.array(emb, dtype=np.float32))
count += 1
if count >= 1000: break
if embeddings:
print(f"✅ Found {len(embeddings)} embeddings for {filename}")
return embeddings
else:
print(f"⚠️ No embeddings found for {filename} in dataset")
return None
except Exception as e:
print(f"❌ Error fetching from HF dataset: {e}")
return None
def analyze_som_region_probability(
som_analyzer: SOMAnalyzer,
center_row: int,
center_col: int,
embeddings: List[np.ndarray],
radius: int = 2
) -> Dict[str, float]:
# (注意:你 main.py 裡的這個函數是同步的(def),不是(async def),這裡保持一致)
print(f" 🔍 Analyzing region around ({center_row}, {center_col}) with radius={radius}")
region_samples = []
for emb in embeddings:
sample_row, sample_col = som_analyzer.som.winner(emb) #
if abs(sample_row - center_row) <= radius and abs(sample_col - center_col) <= radius:
region_samples.append(emb)
print(f" Found {len(region_samples)} samples in region")
if len(region_samples) == 0:
return {"dropper": 0.0, "APT30": 0.0}
feature_counts = {"dropper": [], "spreader": [], "APT30": [], "Codoso_Gh0st_1": []}
if hasattr(som_analyzer, 'background_embeddings') and hasattr(som_analyzer, 'background_labels'):
for i, bg_emb in enumerate(som_analyzer.background_embeddings):
bg_row, bg_col = som_analyzer.som.winner(bg_emb) #
if abs(bg_row - center_row) <= radius and abs(bg_col - center_col) <= radius:
for feature_name in feature_counts.keys():
if feature_name in som_analyzer.background_labels:
feature_counts[feature_name].append(
som_analyzer.background_labels[feature_name][i]
)
probabilities = {}
for feature_name, values in feature_counts.items():
if len(values) > 0:
prob = sum(values) / len(values)
probabilities[feature_name] = float(prob)
else:
probabilities[feature_name] = 0.0
return probabilities
async def analyze_with_som(embedding, predicted_family, filename):
# ... (請從 main.py 完整複製貼上) ...
if not SOM_AVAILABLE:
print("⚠️ SOM analysis disabled")
return None
try:
print("\n🗺️ Performing SOM analysis...")
if isinstance(embedding, list):
embedding_array = np.array(embedding, dtype=np.float32)
else:
embedding_array = embedding
winner_row, winner_col = som_analyzer.find_winner_neuron(embedding_array)
print(f" 📍 Winner neuron: ({winner_row}, {winner_col})")
hf_embeddings = await get_embeddings_from_hf_dataset(filename)
if hf_embeddings and len(hf_embeddings) > 0:
print(f"✅ Using {len(hf_embeddings)} embeddings from HF dataset")
features_prob = analyze_som_region_probability( #
som_analyzer, winner_row, winner_col, hf_embeddings, radius=2
)
data_source = "hf_dataset"
else:
print(f"⚠️ No HF data found, using background data")
features_dict = som_analyzer.predict_features_from_position(
winner_row, winner_col, radius=1
)
features_prob = {k: float(v) for k, v in features_dict.items()}
data_source = "background_data"
features_binary = {
feature: 1 if prob > 0.2 else 0
for feature, prob in features_prob.items()
}
print(f" ✅ Features detected: {features_binary}")
result = {
"winner_position": {"row": int(winner_row), "col": int(col_winner)},
"features": features_binary,
"features_probability": features_prob,
"data_source": data_source
}
print(f"✅ SOM analysis complete!")
return result
except Exception as e:
print(f"❌ SOM analysis error: {e}")
return None
@app.get("/")
def home():
return {
"message": "Malware Classifier API",
"model": MODEL_REPO,
"endpoints": {
"/predict": "POST - Upload TXT files for prediction",
"/model-info": "GET - Get model information"
}
}
@app.get("/model-info")
def get_model_info():
"""取得模型資訊"""
# ✅ 檢查 id2label 的 key 類型
id2label_sample = {}
if hasattr(model.config, 'id2label'):
# 取前 3 個 key 作為範例
sample_keys = list(model.config.id2label.keys())[:3]
id2label_sample = {
k: {
"type": str(type(k).__name__),
"value": model.config.id2label[k]
} for k in sample_keys
}
return {
"model_name": MODEL_REPO,
"num_labels": model.config.num_labels,
"labels": list(model.config.id2label.values()) if hasattr(model.config, 'id2label') else None,
"device": str(DEVICE),
"max_length": 512,
"id2label_sample": id2label_sample # ✅ 除錯用
}
@app.post("/predict")
async def predict_malware(files: List[UploadFile] = File(...)):
"""
批次預測惡意軟體家族
接收多個 TXT 檔案 (來自同一個執行檔的分段)
返回:
- final_label: 多數投票的家族標籤
- embeddings: 最高注意力權重的 embedding
"""
if not files:
raise HTTPException(status_code=400, detail="No files provided")
print(f"\n{'='*60}")
print(f"📥 Received {len(files)} files")
print(f"{'='*60}")
predictions = []
all_embeddings = []
try:
with torch.no_grad():
for idx, file in enumerate(files, 1):
try:
# 讀取檔案內容
content = await file.read()
asm_code = content.decode('utf-8').strip()
if not asm_code:
print(f"⚠️ Skip empty file: {file.filename}")
continue
print(f"📄 [{idx}/{len(files)}] Processing: {file.filename}")
# Tokenize
inputs = tokenizer(
asm_code,
return_tensors="pt",
max_length=512,
truncation=True,
padding=True
)
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
# 模型預測
outputs = model(**inputs)
# 取得預測結果
logits = outputs.logits
probs = torch.nn.functional.softmax(logits, dim=-1)
confidence, predicted_class = torch.max(probs, dim=-1)
# ✅ 取得 label (使用整數 key)
predicted_id = predicted_class.item()
if hasattr(model.config, 'id2label') and predicted_id in model.config.id2label:
label_name = model.config.id2label[predicted_id]
else:
label_name = f"class_{predicted_id}"
print(f" 🔍 Predicted ID: {predicted_id}, Label: {label_name}")
# ✅ 取得 embedding (使用 hidden_states)
if hasattr(outputs, 'hidden_states') and outputs.hidden_states is not None:
# 取最後一層的 [CLS] token embedding
last_hidden = outputs.hidden_states[-1] # (batch, seq_len, hidden_size)
cls_embedding = last_hidden[:, 0, :].cpu().numpy()[0] # (768,)
else:
# 備用方案: 使用 logits
cls_embedding = logits.cpu().numpy()[0]
# ✅ 取得注意力權重
max_attention_score = 0.0
if hasattr(outputs, 'attentions') and outputs.attentions is not None:
# 取最後一層的注意力,平均所有 head
last_attention = outputs.attentions[-1] # (1, num_heads, seq_len, seq_len)
avg_attention = last_attention.mean(dim=1) # (1, seq_len, seq_len)
# 取 [CLS] token 對所有 token 的注意力
cls_attention = avg_attention[0, 0, :].cpu().numpy() # (seq_len,)
max_attention_score = float(cls_attention.max())
# 儲存結果
predictions.append({
"filename": file.filename,
"predicted_label": label_name,
"confidence": float(confidence.item()),
"attention_score": max_attention_score
})
all_embeddings.append({
"filename": file.filename,
"label": label_name,
"embedding": cls_embedding.tolist(),
"attention_score": max_attention_score
})
print(f" ✅ Predicted: {label_name} (confidence: {confidence.item():.3f})")
except Exception as e:
print(f" ❌ Error processing {file.filename}: {e}")
continue
if not predictions:
raise HTTPException(status_code=400, detail="No valid files processed")
# ===== 多數投票 (Majority Voting) =====
label_votes = [p["predicted_label"] for p in predictions]
vote_counts = Counter(label_votes)
final_label = vote_counts.most_common(1)[0][0]
final_count = vote_counts[final_label]
print(f"\n📊 Vote distribution: {dict(vote_counts)}")
print(f"🏆 Final prediction: {final_label} ({final_count}/{len(predictions)} votes)")
# ===== 選擇最高注意力權重的 embedding (屬於最終預測的 label) =====
# 過濾出預測為 final_label 的 embeddings
same_label_embeddings = [
emb for emb in all_embeddings
if emb["label"] == final_label
]
if same_label_embeddings:
# 選擇注意力權重最高的
best_embedding = max(same_label_embeddings, key=lambda x: x["attention_score"])
else:
# 如果沒有,就選所有裡面注意力最高的
best_embedding = max(all_embeddings, key=lambda x: x["attention_score"])
print(f"🎯 Selected embedding from: {best_embedding['filename']}")
print(f" Attention score: {best_embedding['attention_score']:.4f}")
# ===== 返回結果 =====
result = {
"final_label": final_label,
"vote_count": final_count,
"total_segments": len(predictions),
"confidence": sum(p["confidence"] for p in predictions) / len(predictions),
"vote_distribution": dict(vote_counts),
"embedding": {
"values": best_embedding["embedding"],
"source_file": best_embedding["filename"],
"attention_score": best_embedding["attention_score"],
"dimension": len(best_embedding["embedding"])
},
"segment_predictions": predictions
}
# ===== 新增:執行 SOM 分析 =====
som_analysis = None
final_embedding_values = result.get("embedding", {}).get("values")
if SOM_AVAILABLE and final_embedding_values:
print("\n🗺️ Performing SOM analysis on HF Space...")
try:
#
som_analysis = await analyze_with_som(
embedding=final_embedding_values,
predicted_family=result.get("final_label", "Unknown"),
filename=files[0].filename #
)
except Exception as e:
print(f"❌ SOM analysis error on HF Space: {e}")
#
result["som_analysis"] = som_analysis
print(f"\n✅ Prediction with SOM complete!\n")
return result
except Exception as e:
print(f"\n❌ Error in predict_malware: {e}")
import traceback
traceback.print_exc()
raise HTTPException(status_code=500, detail=f"Prediction failed: {str(e)}")
@app.post("/predict-batch")
async def predict_batch(files: List[UploadFile] = File(...)):
"""
簡化版: 只返回 label 和 embedding,不包含詳細資訊
"""
result = await predict_malware(files)
return {
"final_label": result["final_label"],
"embedding": result["embedding"]["values"]
}
# ✅ HF Space 啟動入口
if __name__ == "__main__":
import uvicorn
port = int(os.environ.get("PORT", 7860))
uvicorn.run(app, host="0.0.0.0", port=port)