安卓恶意软件分类 

Drebin样本的百度网盘下载链接我放在下面评论区了，大家自行下载。本次实验收到上一次实验启发（[微软恶意软件分类](https://www.cnblogs.com/iloveacm/p/13813325.html)），并采用了这篇博文的实现代码（[用机器学习检测Android恶意代码](https://blog.csdn.net/qq_33256688/article/details/81434389?utm_medium=distribute.pc_aggpage_search_result.none-task-blog-2~all~first_rank_v2~rank_v25-1-81434389.nonecase&utm_term=%E6%9C%BA%E5%99%A8%E5%AD%A6%E4%B9%A0%E6%A3%80%E6%B5%8B%E5%AE%89%E5%8D%93%E6%81%B6%E6%84%8F%E8%BD%AF%E4%BB%B6&spm=1000.2123.3001.4430)），代码都可以在博主提供的github地址中找到。

原理

​ 具体原理参考上一次实验，都是提取反编译文件中的操作码，n-gram中n取3。具体原理可以参考这篇文章。与上次实验不同的是，本次实验针对的是安卓软件，所以具体的操作码有所不同。并且由于所采用的数据集中良性软件明显比恶意软件大的多。所以n-gram不再采用出现频次而是是否出现作为特征。

数据集

​ 本次实验的恶意软件数据集来自于Drebin，只采用了第一个part共1000个恶意软件。良性软件来自于这个网站。共1100多个良性软件，取其中的1000个。良性软件集12.3GB，恶意软件集1.2GB。可以看出良性软件要比恶意软件大的多。

反编译

​ 将良性数据集以及恶意数据集软件分别反编译到 \smalis\kind 以及 \smalis\malware 中。代码如下

1# -*- coding: utf-8 -*-
"""
Created on Tue Feb  6 14:00:51 2018
@author: 燃烧杯
"""

import os
import subprocess

def disassemble(frompath, topath, num, start=0):
    files = os.listdir(frompath)
    files = files[start:num]
        
    total = len(files)
    
    for i, file in enumerate(files):
        fullFrompath = os.path.join(frompath, file)
        fullTopath = os.path.join(topath, file)
        command = "apktool d " + fullFrompath + " -o " + fullTopath
        subprocess.call(command, shell=True)
        print("已反汇编", i, "个应用，百分比如下：")
        print((i + 1) * 100 / total, "%")


#反汇编恶意软件样本
virus_root = "..\\bit\\virus\\VirusAndroid"
disassemble(virus_root, ".\\smalis\\malware", 600)


#反汇编正常软件样本
kind_root = "..\\bit\\virus\\normalApk"
disassemble(kind_root, ".\\smalis\\kind", 600)

​ 完成后每个软件会创建一个以文件名字命名的文件夹，文件夹中包含反编译后的文件，如下图所示：

image-20201015172056064

​ 其中smali文件夹中包含了我们要提取特征码的文件。smali文件大致如下

image-20201015172738379

​ 我们要提取的操作码就在.method中。操作码大概有下图这几类

p1

​ 将每一类的操作码对应为大写字母以简化特征码。如move表示为M。

操作码提取

​ 代码对应上文提到的github中的bytecode_extract.py文件。

# -*- coding: utf-8 -*-
"""
Created on Tue Feb  6 22:41:06 2018
@author: 燃烧杯
"""

from infrastructure.ware import Ware
from infrastructure.fileutils import DataFile

virusroot = "./smalis/malware"
kindroot = "./smalis/kind"

f = DataFile("./data.csv")

import os

def collect(rootdir, isMalware):
    wares = os.listdir(rootdir)
    total = len(wares)
    for i, ware in enumerate(wares):
        warePath = os.path.join(rootdir, ware)
        ware = Ware(warePath, isMalware)
        ware.extractFeature(f)
        print("已提取", i + 1, "个文件的特征，百分比如下：")
        print((i + 1) * 100 / total, "%")
        
    
#1代表恶意软件
collect(virusroot, 1)
collect(kindroot, 0)    

f.close()

​ 提取出后如下图所示：

image-20201015172554235

​ feture列就是我们为每个文件提取出的特征。每个方法的特征码序列用“|”隔开。

n-gram特征

​ 从上文的feture中提取出n-gram特征，其数值表示该操作序列是否出现。代码如下

# -*- coding: utf-8 -*-
"""
Created on Fri Feb  9 13:26:50 2018
@author: 燃烧杯
词集模型
"""

import sys

#n-gram的n值
n = int(sys.argv[1])
print("n = ", n)

import pandas as pd

origin = pd.read_csv("data.csv")
#origin = pd.read_csv("test.csv")

from infrastructure.mydict import MyDict

mdict = MyDict()

feature = origin["Feature"].str.split("|")
total = len(feature)
for i, code in enumerate(feature):
    mdict.newLayer()
    if not type(code) == list:
        continue
    for method in code:
        length = len(method)
        if length < n:
            continue
        for start in range(length - (n - 1)):
            end = start + n
            mdict.mark(method[start:end])
    print("已完成", i, "个应用，百分比如下：")
    print((i + 1) * 100 / total, "%")
            
result = mdict.dict
pd.DataFrame(result, index=origin.index)\
               .to_csv("./" + str(n) + "_gram.csv", index=False)

​ 结果如图：

image-20201015173708451

形成了2000343的特征表，之所以是343个特征序列是应为总共有7大类操作码，并且采用3-gram，有777个序列。

机器学习

​ 接下来就是训练了，本次实验采用随机森林算法，并采用10交叉验证，代码如下：

from sklearn.ensemble import RandomForestClassifier as RF
from sklearn.model_selection import cross_val_score
from sklearn.metrics import confusion_matrix
import pandas as pd

train_feture = pd.read_csv('D:\\android\\dataset\\smalis\\3_gram.csv')
data = pd.read_csv('D:\\android\\dataset\\smalis\\data.csv')
labels = data["isMalware"]
train_feture = train_feture.iloc[:,:].values
srf = RF(n_estimators=500, n_jobs=-1)
clf_s = cross_val_score(srf, train_feture, labels, cv=10)

​ 结果如下

array([0.965     , 0.995     , 0.995     , 0.96      , 0.89      ,
       0.945     , 0.965     , 0.95      , 0.97487437, 0.97487437])

深度学习

​ 顺便用用深度学习做一下分类看看效果，深度学习库采用keras。以下是代码：

from tensorflow import keras
from tensorflow.keras import layers
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
%matplotlib inline
import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"

test_split = 0.2			#划分训练集与测试集
data = pd.read_csv("D:\\android\\dataset\\smalis\\data_2.csv")
fetrues = pd.read_csv("D:\\android\\dataset\\smalis\\3_gram.csv")
labels = data["isMalware"]

p1 = int(len(labels)*(1-test_split))
index = np.random.permutation(len(fetrues))		#打乱顺序
train_data = fetrues.iloc[index]
labels = labels.iloc[index]
index = np.random.permutation(len(fetrues))
train_data = fetrues.iloc[index]
labels = labels.iloc[index]

model = keras.Sequential()
model.add(layers.Dense(50,input_dim = 343, activation = 'relu'))
model.add(layers.Dense(16, activation = 'relu'))
model.add(layers.Dense(16, activation = 'relu'))
model.add(layers.Dense(16, activation = 'relu'))
model.add(layers.Dense(16, activation = 'relu'))
model.add(layers.Dense(16, activation = 'relu'))
model.add(layers.Dense(1, activation = 'sigmoid'))
model.compile(
    optimizer = 'adam',
    loss='binary_crossentropy',
    metrics=['acc']
)

history = model.fit(x_train, y_train, epochs=60, batch_size=256, validation_data=(x_test, y_test))

​ 测试结果如下：

​ image-20201016093540843

最后10轮精确度如下：

0.9812, 0.9819, 0.9775, 0.9781, 0.9718, 0.9812, 0.9793, 0.9618, 0.9825, 0.9756

另外做10交叉验证，代码如下：

from tensorflow import keras
from tensorflow.keras import layers
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
%matplotlib inline
import os

os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"

from sklearn.model_selection import StratifiedKFold
seed = 7
np.random.seed(seed)
kfold = StratifiedKFold(n_splits=10, shuffle=True, random_state=seed)

cvscores = []
data = pd.read_csv("D:\\android\\dataset\\smalis\\data_2.csv")
labels = data["isMalware"]
train_data = pd.read_csv("D:\\android\\dataset\\smalis\\3_gram.csv")
train_data =  train_data.iloc[:,:].values
for train, test in kfold.split(train_data, labels):
    model = keras.Sequential()
    model.add(layers.Dense(50,input_dim = 343, activation = 'relu'))
    model.add(layers.Dense(16, activation = 'relu'))
    model.add(layers.Dense(16, activation = 'relu'))
    model.add(layers.Dense(16, activation = 'relu'))
    model.add(layers.Dense(16, activation = 'relu'))
    model.add(layers.Dense(16, activation = 'relu'))
    model.add(layers.Dense(1, activation = 'sigmoid'))
    model.compile(
    optimizer = 'adam',
    loss='binary_crossentropy',
    metrics=['acc']
    )
    model.fit(train_data[train],labels[train],epochs=60, batch_size=256,verbose = 0)
    scores = model.evaluate(train_data[test], labels[test], verbose=0)
    print(scores[1])
    cvscores.append(scores[1])
print(cvscores)

精确度如下：

[0.95, 0.985, 0.95, 0.945, 0.975, 0.95, 0.955, 0.96, 0.9798995, 0.9748744]

与随机森林对比图;

image-20201016102523772

posted @ 2020-10-15 17:45  iloveacm  阅读(4760)  评论(69)  编辑  收藏  举报