ma_cisco_malware/dataset.py

# -*- coding: utf-8 -*-
import string

import h5py
import numpy as np
import pandas as pd
from keras.utils import np_utils
from tqdm import tqdm

chars = dict((char, idx + 1) for (idx, char) in
             enumerate(string.ascii_lowercase + string.punctuation + string.digits))


def get_character_dict():
    return chars


def encode_char(c):
    if c in chars:
        return chars[c]
    else:
        return 0


encode_char = np.vectorize(encode_char)


# TODO: refactor
def get_user_chunks(dataFrame, windowSize=10, overlapping=False,
                    maxLengthInSeconds=300):
    maxMilliSeconds = maxLengthInSeconds * 1000
    outDomainLists = []
    outDFFrames = []
    if not overlapping:
        numBlocks = int(np.ceil(float(len(dataFrame)) / float(windowSize)))
        userIDs = np.arange(len(dataFrame))
        for blockID in np.arange(numBlocks):
            curIDs = userIDs[(blockID * windowSize):((blockID + 1) * windowSize)]
            # print(curIDs)
            useData = dataFrame.iloc[curIDs]
            curDomains = useData['domain']
            if maxLengthInSeconds != -1:
                curMinMilliSeconds = np.min(useData['timeStamp']) + maxMilliSeconds
                underTimeOutIDs = np.where(np.array(useData['timeStamp']) <= curMinMilliSeconds)
                if len(underTimeOutIDs) != len(curIDs):
                    curIDs = curIDs[underTimeOutIDs]
                    useData = dataFrame.iloc[curIDs]
                    curDomains = useData['domain']
            outDomainLists.append(list(curDomains))
            outDFFrames.append(useData)
    else:
        numBlocks = len(dataFrame) + 1 - windowSize
        userIDs = np.arange(len(dataFrame))
        for blockID in np.arange(numBlocks):
            curIDs = userIDs[blockID:blockID + windowSize]
            useData = dataFrame.iloc[curIDs]
            curDomains = useData['domain']
            if maxLengthInSeconds != -1:
                curMinMilliSeconds = np.min(useData['timeStamp']) + maxMilliSeconds
                underTimeOutIDs = np.where(np.array(useData['timeStamp']) <= curMinMilliSeconds)
                if len(underTimeOutIDs) != len(curIDs):
                    curIDs = curIDs[underTimeOutIDs]
                    useData = dataFrame.iloc[curIDs]
                    curDomains = useData['domain']
            outDomainLists.append(list(curDomains))
            outDFFrames.append(useData)
    if len(outDomainLists[-1]) != windowSize:
        outDomainLists.pop(-1)
        outDFFrames.pop(-1)
    return outDomainLists, outDFFrames


def get_domain_features(domain, vocab: dict, max_length=40):
    encoding = np.zeros((max_length,))
    for j in range(min(len(domain), max_length)):
        char = domain[-j]  # TODO: why -j -> order reversed for domain url?
        encoding[j] = vocab.get(char, 0)
    return encoding


def get_all_flow_features(features):
    flows = np.stack(list(
        map(lambda f: f[["duration", "bytes_up", "bytes_down"]], features))
    )
    return np.log1p(flows)


def create_dataset_from_flows(user_flow_df, char_dict, max_len, window_size=10):
    domains = []
    features = []
    print("get chunks from user data frames")
    for i, user_flow in tqdm(list(enumerate(get_flow_per_user(user_flow_df)))):
        (domain_windows, feature_windows) = get_user_chunks(user_flow,
                                                            windowSize=window_size,
                                                            overlapping=False,
                                                            maxLengthInSeconds=-1)
        domains += domain_windows
        features += feature_windows

    print("create training dataset")
    domain_tr, flow_tr, hits_tr, _, server_tr, trusted_hits_tr = create_dataset_from_lists(domains=domains,
                                                                                           flows=features,
                                                                                           vocab=char_dict,
                                                                                           max_len=max_len,
                                                                                           window_size=window_size)

    # make client labels discrete with 4 different values
    hits_tr = np.apply_along_axis(lambda x: discretize_label(x, 3), 0, np.atleast_2d(hits_tr))
    # select only 1.0 and 0.0 from training data
    pos_idx = np.where(np.logical_or(hits_tr == 1.0, trusted_hits_tr >= 1.0))[0]
    neg_idx = np.where(hits_tr == 0.0)[0]
    idx = np.concatenate((pos_idx, neg_idx))
    # choose selected sample to train on
    domain_tr = domain_tr[idx]
    flow_tr = flow_tr[idx]
    client_tr = np.zeros_like(idx, float)
    client_tr[:pos_idx.shape[-1]] = 1.0
    server_tr = server_tr[idx]

    client_tr = np_utils.to_categorical(client_tr, 2)
    server_tr = np_utils.to_categorical(server_tr, 2)

    return domain_tr, flow_tr, client_tr, server_tr


def store_h5dataset(path, domain_tr, flow_tr, client_tr, server_tr):
    f = h5py.File(path, "w")
    domain_tr = domain_tr.astype(np.int8)
    f.create_dataset("domain", data=domain_tr)
    f.create_dataset("flow", data=flow_tr)
    server_tr = server_tr.astype(np.bool)
    client_tr = client_tr.astype(np.bool)
    f.create_dataset("client", data=client_tr)
    f.create_dataset("server", data=server_tr)
    f.close()


def load_h5dataset(path):
    data = h5py.File(path, "r")
    return data["domain"], data["flow"], data["client"], data["server"]


def create_dataset_from_lists(domains, flows, vocab, max_len, window_size=10):
    """
    combines domain and feature windows to sequential training data
    :param domains: list of domain windows
    :param flows: list of flow feature windows
    :param vocab: 
    :param max_len: 
    :param window_size: size of the flow window
    :return: 
    """
    # sample_size = len(domains)

    # domain_features = np.zeros((sample_size, window_size, max_len))
    flow_features = get_all_flow_features(flows)

    domain_features = np.array([[get_domain_features(d, vocab, max_len) for d in x] for x in domains])

    hits = np.max(np.stack(map(lambda f: f.virusTotalHits, flows)), axis=1)
    names = np.unique(np.stack(map(lambda f: f.user_hash, flows)), axis=1)
    servers = np.max(np.stack(map(lambda f: f.serverLabel, flows)), axis=1)
    trusted_hits = np.max(np.stack(map(lambda f: f.trustedHits, flows)), axis=1)

    return (domain_features, flow_features,
            hits, names, servers, trusted_hits)


def discretize_label(values, threshold):
    maxVal = np.max(values)
    if maxVal >= threshold:
        return 1.0
    elif maxVal == -1:
        return -1.0
    elif 0 < maxVal < threshold:
        return -2.0
    else:
        return 0.0


def get_user_flow_data(csv_file):
    types = {
        "duration": int,
        "bytes_down": int,
        "bytes_up": int,
        "domain": object,
        "timeStamp": float,
        "server_ip": object,
        "user_hash": float,
        "virusTotalHits": int,
        "serverLabel": int,
        "trustedHits": int
    }
    df = pd.read_csv(csv_file)
    df = df[list(types.keys())]
    df.set_index(keys=['user_hash'], drop=False, inplace=True)
    return df


def get_flow_per_user(df):
    users = df['user_hash'].unique().tolist()
    for user in users:
        yield df.loc[df.user_hash == user]