PREDICTING ZERO-DAY SOFTWARE VULNERABILITIES THROUGH DATA MINING --SECOND PRESENTATION Su Zhang 1
Feb 25, 2016
1
PREDICTING ZERO-DAY SOFTWARE VULNERABILITIES THROUGH DATA
MINING--SECOND PRESENTATION
Su Zhang
2
Quick Review. Data Source – NVD. Six Most Popular/Vulnerable Vendors For Our
Experiments. Why The Six Vendors Are Chosen. Data Preprocessing. Functions Available For Our Approach. Statistical Results Plan For Next Phase.
Outline
3
Quick Review
4
National Vulnerability Database◦ U.S. government repository of standards based
vulnerability management data.◦ Data included in each NVD entry
Published Date Time Vulnerable software’s CPE Specification
◦ Derived data Published Date Time Month Published Date Time Day Two adjacent vulnerabilities’ CPE diff (v1,v2)Version diff CPE Specification Software Name Adjacent different Published Date Time ttpv Adjacent different Published Date Time ttnv
Source Database – NVD
5
Linux: 56925 instances Sun: 24726 instances Cisco: 20120 instances Mozilla: 19965 instances Microsoft: 16703 instances Apple: 14809 instances.
Six Most Vulnerable/Popular Vendors
6
r e s tAd
obe IBM Ph
pAp
ple
Microso
ft
Mozilla
Cisco
SunLin
ux0
100002000030000400005000060000
Instances Table
Instances
Why We Only Choose Instances Of Pop Vendors—Instances Table
7
r e s t HPLin
uxMozi
laCisc
oOrac
le IBMApple Su
n
Microso
ft0
500
1000
1500
2000
2500Vulnerability Table
Vul_Num
Why We Only Choose Instances Of Pop Vendors—Vulnerability Table
8
Huge size of nominal types (vendors and software) will result in a scalability issue.
Top six take up 43.4% of all instances.
We have too many vendors(10411) in NVD.
The seventh most popular/vulnerable vendor is much less than the sixth.
Vendors are independent for our approach.
Why We Only Choose Instances Of Pop Vendors
9
NVD data—Training/Testing dataset◦ Starting from 2005 since before that the data
looks unstable.◦ Correct some obvious errors in NVD(e.g.
“cpe:/o:linux:linux_kernel:390”).
Attributes◦ Published time : Only use month and day. ◦ Version diff: A normalized difference between two
versions.◦ Vendor: Removed.
Data Preprocessing
10
Attributes◦ “Group” vulnerabilities published at the same
day- we can guarantee ttnv/ttpv are non-zero values.
◦ ttnv is the predicted attribute.
For each software◦ Delete its first bunch of instances.◦ Delete its last bunch of instances.
Data Preprocessing(cont)
11
v1= 3.6.4; v2 = 3.6; MaxVersionLength=4; v1= expand ( v1, 4 ) = 3.6.4.0 v2 =expand ( v2, 4 ) = 3.6.0.0 diff(v1, v2) = (3-3) * 1000 +(6-6) * 100-1
+(4-0) * 100-2
+(0-0) * 100-3 = 4 E -4
version diff Calculation
12
Vendor, soft, version, month, day, vdiff, ttpv, ttnv linux,kernel,2.6.18, 05, 02, 0, 70, 5 linux,kernel,2.6.19.2, 05, 07,1.02E-4,5, 281
An Example
13
Least Mean Square. Linear Regression Multilayer Perceptron. SMOreg. RBF Network. Gaussian Processes.
Functions Available For Our Approach On Weka
14
Function: Linear Regression Training Dataset: 66% Linux(Randomly picked
since 2005). Test Dataset: the rest 34% Test Result:
◦ Correlation coefficient 0.5127◦ Mean absolute error 11.2358◦ Root mean squared error 25.4037◦ Relative absolute error 107.629 %◦ Root relative squared error 86.0388 %◦ Total Number of Instances 17967
Several Statistical Results
15
Correlation Coefficient
16
Mean absolute error :
Root mean square error:
Several Definitions About “Error”
17
Relative absolute error:
Root relative squared error:
Several Definitions About “Error”(Cont)
18
Function: Least Mean Square Training Dataset: 66% Linux(Randomly picked
since 2005). Test Dataset: the rest 34% Test Result:
◦ Correlation coefficient -0.1501◦ Mean absolute error 7.6676◦ Root mean squared error 30.6038◦ Relative absolute error 73.449 %◦ Root relative squared error 103.6507 %◦ Total Number of Instances 17967
Several Statistical Results
19
Function: Multilayer Perceptron Training Dataset: 66% Linux(Randomly picked
since 2005). Test Dataset: the rest 34% Test Result:
◦ Correlation coefficient 0.9886◦ Mean absolute error 0.4068◦ Root mean squared error 4.6905◦ Relative absolute error 3.7802 %◦ Root relative squared error 15.1644 %◦ Total Number of Instances 17967
Several Statistical Results
20
Function: RBF Network Training Dataset: 66% Linux(Randomly picked since
2005). Test Dataset: the rest 34% Test Result:
◦ Linear Regression Model ttnv = -15.3206 * pCluster_0_1 + 21.6205
◦ Correlation coefficient 0.1822◦ Mean absolute error 10.5857◦ Root mean squared error 29.048 ◦ Relative absolute error 101.4023 %◦ Root relative squared error 98.3814 %◦ Total Number of Instances 17967
Several Statistical Results
21
Linear Regression: Not accurate enough but looks promising (correlation coefficient: 0.5127).
Least Mean Square: Probably not good for our approach(negative correlation coefficient).
Multilayer Perceptron: Looks good but it couldn’t provide us with a linear model.
Summary Of Current Results
22
SMOreg: For most vendors, it takes too long time to finish (usually more than 80 hours).
RBF Network: Not very accurate.
Gaussian Processes: Runs out of heap memory for most of our experiments.
Summary Of Current Results (Cont)
23
Adding CVSS metrics as predictive attributes.
Binarize our predictive attributes (e.g. divide ttnv/ttpv into several categories.)
Use regression SVM with multiple kernels.
Possible Ways To Improve The Accuracy Of Our Models.
24
Try to find out an optimal model for our prediction.
Try to investigate how to apply it with MulVAL if we get a good model. Otherwise, find out the reason why it is not accurate enough.
Plan For Next Phase
25
Thank you!