第二届高校大数据比赛之鼠标轨迹识别

比赛地址http://bdc.saikr.com/c/cql/34541

赛题

鼠标轨迹识别当前广泛运用于多种人机验证产品中，不仅便于用户的理解记忆，而且极大增加了暴力破解难度。但攻击者可通过黑产工具产生类人轨迹批量操作以绕过检测，并在对抗过程中不断升级其伪造数据以持续绕过同样升级的检测技术。我们期望用机器学习算法来提高人机验证中各种机器行为的检出率，其中包括对抗过程中出现的新的攻击手段的检测。

数据格式

评测指标

F = 5PR/(2P+3R)*100

数据读取和处理

---

#####数据读取和处理
import pandas as pd
import os


def get_data(file):
    data1=[]
    count=0
    with open(file) as f:
        for i in f.readlines():
            count+=1
            arr=i.split(" ")[1].split(";")[:-1]
            for j in arr:
                temp=[count]
                temp.extend(j.split(","))
                data1.append(temp)
    data2=[]
    with open(file) as f:
        for i in f.readlines():
            count+=1
            arr=i.split(" ")[2]
            data2.append(arr.split(","))

    data=pd.DataFrame(data1,columns=["id","x","y","t"])
    d2=pd.DataFrame(data2,columns=["target_x","target_y"])
    d2.target_y=d2.target_y.apply(lambda x:x[:-1])
    d2["id"]=range(1,100001)
    data=pd.merge(data,d2,on="id")
    return data

数据可视化

---

import matplotlib.pyplot as plt
%matplotlib inline
# plt.xticks(list(range(len(b))),b["x"].values)
import os 
path="F:\competition_data\Bigdata\images"
# os.mkdir(path)
for i in range(1,3001):
    b=data[data.id==i]
    k=list(b["x"].values)
#     k.extend(set(b["target_x"].values))
    l=list(b["y"].values)
#     l.extend(set(b["target_y"].values))
    plt.plot(k,l,"o-")
    fig = plt.gcf()
    fig.set_size_inches(30, 15)
    fig.savefig(path+"\"+str(i)+".png",dpi=100)
    plt.close()

特征提取

---

###特征提取
def get_features(data):
    a=pd.DataFrame()
    data_length=len(set(data.id.values))
    import numpy as np
    for i in range(data_length):
        test=data[data.id==i]
        if len(test)!=1:
            test.index=range(len(test))
            temp=test[["x","y","t"]].diff(1).dropna()
            temp["distance"]=np.sqrt(temp["x"]**2+temp["y"]**2)
            temp["speed"]=np.log1p(temp["distance"])-np.log1p(temp["t"])
            temp["angles"]=np.log1p(temp["y"])-np.log1p(temp["x"])
            speed_diff=temp["speed"].diff(1).dropna()
            angle_diff=temp["angles"].diff(1).dropna()
            test["distance_aim_deltas"]=np.sqrt((test["x"]-test["target_x"])**2+(test["y"]-test["target_y"])**2)
            distance_aim_deltas_diff=test["distance_aim_deltas"].diff(1).dropna()

            arr=pd.DataFrame(index=[0])
            arr["id"]=i
            arr["speed_diff_median"] = speed_diff.median()
            arr["speed_diff_mean"] = speed_diff.mean()
            arr["speed_diff_var"] =  speed_diff.var()
            arr["speed_diff_max"] = speed_diff.max()
            arr["angle_diff_var"] =  angle_diff.var()
            arr["time_delta_min"] =  temp["t"].min()
            arr["time_delta_max"] = temp["t"].max()
            arr["time_delta_var"] = temp["t"].var()

            arr["distance_deltas_max"] =  temp["distance"].max()
            arr["distance_deltas_var"] =  temp["distance"].var()
            arr["aim_distance_last"] = test["distance_aim_deltas"].values[-1]
            arr["aim_distance_diff_max"] = distance_aim_deltas_diff.max()
            arr["aim_distance_diff_var"] = distance_aim_deltas_diff.var()
            arr["mean_speed"] = temp["speed"].mean()
            arr["median_speed"] = temp["speed"].median()
            arr["var_speed"] = temp["speed"].var()

            arr["max_angle"] = temp["angles"].max()
            arr["var_angle"] =  temp["angles"].var()
            arr["kurt_angle"] =  temp["angles"].kurt()

            arr["y_min"] = test["y"].min()
            arr["y_max"] = test["y"].max()
            arr["y_var"] = test["y"].var()
            arr["y_mean"] = test["y"].mean()
            arr["x_min"] = test["x"].min()
            arr["x_max"] = test["x"].max()
            arr["x_var"] = test["x"].var()
            arr["x_mean"] = test["x"].mean()

            arr["x_back_num"] = min( (test["x"].diff(1).dropna() > 0).sum(), (test["x"].diff(1).dropna() < 0).sum())
            arr["y_back_num"] = min( (test["y"].diff(1).dropna() > 0).sum(), (test["y"].diff(1).dropna() < 0).sum())

            arr["xs_delta_var"] = test["x"].diff(1).dropna().var()
            arr["xs_delta_max"] = test["x"].diff(1).dropna().max()
            arr["xs_delta_min"] =test["x"].diff(1).dropna().min()
    #         arr["label"]=test["label"]
            a=pd.concat([a,arr])
        return a

模型

---

###xgb
import xgboost as xgb
test_x=test.drop("id",1)
train_x=train.drop(["id","label"],1)

dtest = xgb.DMatrix(test_x)
# dval = xgb.DMatrix(val_x,label=val_data.label)
dtrain = xgb.DMatrix(train_x, label=train.label)
params={
    "booster":"gbtree",
    "objective": "binary:logistic",

#   "scale_pos_weight": 1500.0/13458.0,
        "eval_metric": "auc",

    "gamma":0.1,#0.2 is ok
    "max_depth":3,
#   "lambda":550,
        "subsample":0.7,
        "colsample_bytree":0.4 ,
#         "min_child_weight":2.5, 
        "eta": 0.007,
#     "learning_rate":0.01,
    "seed":1024,
    "nthread":7,

    }

watchlist  = [(dtrain,"train"),
# (dval,"val")
             ]#The early stopping is based on last set in the evallist
model = xgb.train(
    params,
                  dtrain,
                  feval=feval,
#                   maximize=False,

                          num_boost_round=1500,
#                   early_stopping_rounds=10,
#                   verbose_eval =30,
                  evals=watchlist
                 )
# model=xgb.XGBClassifier( 
# max_depth=4,
#     learning_rate=0.007, 
#     n_estimators=1500,
#     silent=True,
#     objective="binary:logistic",
# #     booster="gbtree",
# #     n_jobs=-1, 
#     nthread=7, 
# #     gamma=0, 
# #     min_child_weight=1,
# #     max_delta_step=0,
#     subsample=0.7, 
#     colsample_bytree=0.7, 
# #     colsample_bylevel=0.7,
# #     reg_alpha=0,
# #     reg_lambda=1, 
#     scale_pos_weight=1,
#     base_score=0.5,
# #     random_state=0,
#     seed=1024,
#     missing=None, 
# )

# xgb.cv(params,dtrain,num_boost_round=1500,nfold=10,feval=feval,early_stopping_rounds=50,)
# model.save_model("./model/xgb.model")
# print "best best_ntree_limit",model.best_ntree_limit  

评价函数

---

def eval(clf,x,y):
    prob=clf.predict(x)
    for i in range(len(prob)):
        if prob[i]>=1:
            prob[i]=1
        else:
            prob[i]=0
    p=((y==0)&(prob==0)).sum()/(prob==0).sum()
    print("TP"+" : "+str(((y==0)&(prob==0)).sum())+"  "+"预测"+" : "+str((prob==0).sum())+"  "+"真实"+" : "+str((y==0).sum()))
    r=((y==0)&(prob==0)).sum()/(y==0).sum()
    if p==0 or r==0:
        print(0.0)
        return 0.0

    f=5*p*r/(2*p+3*r)*100
    print(f)
    return f
def feval(pred,dtrain):
    y=dtrain.get_label()
    for i in range(len(pred)):
        if pred[i]>=0.5:
            pred[i]=1
        else:
            pred[i]=0
    p=((y==0)&(pred==0)).sum()/(pred==0).sum()
    print("---------------------------------------------------------")
#     print("TP"+" : "+str(((y==0)&(pred==0)).sum())+"  "+"预测"+" : "+str((pred==0).sum())+"  "+"真实"+" : "+str((y==0).sum()))
    r=((y==0)&(pred==0)).sum()/(y==0).sum()
    if p==0 or r==0:
        print(0.0)
        return "f",0.0

    f=5*p*r/(2*p+3*r)*100
    print(f)
    return "f",f
def target(score,num):
    x=score*(40000+3*num)/5
    return x

线下cv

---

from sklearn import cross_validation
score=cross_validation.cross_val_score(m,train.ix[:,1:-1],train.label,cv=10,scoring=eval)
score.mean()

提交结果

---

pred=model.predict(dtest)
test["prob"]=pred
submit=test.sort_values(by="prob").head(20000)
submit=submit[["id"]]
submit=submit.astype(int)

线上成绩0.91