differential testing - cis.upenn.edumhnaik/edu/cis700/lessons/differential... · Mayur Naik CIS 700 –Fall 2018 Motivation Providing functional specificationsfor large software is

9/24/18

1

DifferentialTesting

MayurNaikCIS700– Fall2018

Motivation

Providingfunctionalspecifications forlargesoftwareisveryhard

• Functionalspecification:desiredpropertiesofinputsandoutputs

• E.g.,Ccompiler

> 600pagesinEnglish

=?

C Compiler

Input

Output

9/24/18

2

AnIdea

C Compiler

Input

Other C Compilers

Compare OutputBuggy Correct

different same

DifferentialTesting

Cross-check– Executedifferentimplementationsofthesamefunctionality(e.g.,GCC andLLVM)withthesameinputs

– Comparetheiroutputs– Report anyanomaliesasbugs

9/24/18

3

SuccessStories

• Compilers

• NeuralNets

int foo() {signed char x = 1;unsigned char y = 255;return x > y;

}

1

gcc inUbuntu8.04

Othercompilers

0Buggy Correct

OneversionofNvidia self-drivingcar

systemOtherversions

Correct“Go straight” “Turn left”

Buggy

Challenges

Howdowegenerategoodinputs?

– Concise:Avoidsillegalandredundanttests

– Diverse:Givesgoodcoverageofdiscrepantparts

9/24/18

4

Approaches

• UnguidedApproach– Generatetestinputsindependentlyacrossiterations

– e.g.,Csmith• GuidedApproach

– Generatetestinputsbyobservingprogrambehaviorforpastinputs

– e.g.,NEZHA,DeepXplore

UnguidedTestingforCCompilers

CSmith:UnguidedDifferential Testing toolforCCompilers

Random C program generator

C-program

Binary 1

...

Binary 2

Binary n

Compare

9/24/18

5

InputGeneration

• Basedonrandomtesting– RandomlygenerateCprograms

• Consideringdomain-specificknowledge– Well-formedness(Csyntax)– Well-definedness (Csemantics)

FoundCompilerBugs

LLVMVersion

1.9 2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8

#Bugs 27 20 18 22 22 12 13 7 10 1

GCCVersion

3.0.0 3.1.0 3.2.0 3.3.0 3.4.0 4.0.0 4.1.0 4.2.0 4.3.0 4.4.0

#Bugs 10 11 9 7 7 11 7 6 14 5

9/24/18

6

LimitationsofUnguidedTesting

Generally,highlyinefficienttofinddiscrepancies,because:

– Randomlygenerates inputs

– Ignores anyinformationfrompastinputs

GuidedTestingforBinaries

NEZHA:GuidedDifferential Testing tool forBinaries

– Exploitbehavioraldiscrepanciesbetweenmultiplebinaryexecutables

– Evolveaninputcorpusthatisguidedbasedonruntimeinformation(obtainedbydoingbinaryinstrumentation)

9/24/18

7

GuidedTesting

GuidedTesting

• Startingfrominputseeds,keepgeneratingmutationsuntildiscrepanciesarefound

• Insteadofmaintainingallmutants,onlykeep“promising”ones– Mutantsofinputsthatledtocombinations

ofstatesunseensofar

9/24/18

8

Example

Inputcorpus:{ 7,0,1}

Exercisedinput:

Exercisededges: { }

Example

Inputcorpus:{ 7,0,1}

Exercisedinput:7

Exercisededges:{

{E1,E2’,E3’}}

9/24/18

9

Example

Inputcorpus:{ 7, 0,1 }

Exercisedinput:0

Exercisededges:{

{E1,E2’,E3’},{E2,E3,E1’}

}

Example

Inputcorpus:{ 7, 0,1 }

Exercisedinput:1

Exercisededges:{

{E1,E2’,E3’},{E2,E3,E1’},{E1,E1’}

}

Notaddedanddiscarded

9/24/18

10

Example

Inputcorpus:{ 7, 0,1 }

Exercisedinput:2(pick1andmutateitbyadding1)

Exercisededges:{

{E1,E2’,E3’},{E2,E3,E1’},{E1,E1’},{E3,E4,E1’}

}

Intuition

• Inputsthatexercisedifferentcoderegions

inthetwoappsmightimplydifferencesin

handlinglogic

• Suchinputsarelikelytofinddiscrepancies

9/24/18

11

CaseStudies

Application Tests

SSLLibraries OpenSSL,LibreSSL,BoringSSL,GnuTLS,wolfSSL,mbedTLS

PDFReaders EvincePDF,MuPDF,Xpdf

ELF Parser ClamAV,binutils

XZParser ClamAV,XZ

DiscrepanciesandBugsFound

Type SSLCertificate

PDFFile

ELFBinary

XZArchive

Discrepancies 764 7 2 5

ErrorsandCrashes 6 0 0 2

9/24/18

12

GuidedTestingforDNNs

DeepXplore:guideddifferentialtestingtoolforDeepNeuralNetworks(DNNs)

GuidedTestingforDNNs

Decision boundary of DNN 1 Decision boundary

of DNN 2

Seed input

9/24/18

13

InputGeneration

• Iteratively changing a seed input• Two criteria:1) Maximizing differential behaviors2) Maximizing neuron coverage

• Formally, maximizing the following objective function

obj(x) = (𝚺k≠jF(x)[c] - ƛ1∙Fj(x)[c]) + ƛ2∙fn(x)

obj1 obj2

MaximizingDifferentialBehavior

• Supposewehaven DNNsF1…Fn• LetFk(x)[c]betheclassprobability thatFk(x)predicts xtobe c

• RandomlyselectoneDNNFk• Maximizethisobjectivefunction:

• Easilysolvedusinggradientascent

obj1(x) = 𝚺k≠jF(x)[c] - ƛ1∙Fj(x)[c]

9/24/18

14

MaximizingNeuronCoverage

• Iterativelypickaninactivatedneuronf

• Modifytheinputsuchthataneuronbecomesactivated,thatis,maximizethis objectivefunctionforneuronn:

• Easilysolvedusinggradientascent

obj2(x) = fn(x)

EmpiricalEvaluation

Dataset Description DNN Description DNN Name #Neurons #Discripancies/ #Seeds

MNIST Hand-written digits LeNet variations

MNI_C1MNI_C2MNI_C3

52148268

1,073 / 2,0001,968 / 2,000

827 / 2,000

Imagenet General images

State-of-the-art image classifiers

IMG_C1IMG_C2IMG_C3

14,88816,16894,059

1,969 / 2,0001,976 / 2,0001,996 / 2,000

Driving Driving video frames

Nvidia DAVE self-driving systems

DRV_C1DRV_C2DRV_C3

1,5601,560

844

1,720 / 2,0001,866 / 2,0001,930 / 2,000

Contagio/Virustoral PDFs PDF malware detectors

PDF_C1PDF_C2PDF_C3

402602802

1,103 / 2,000789 / 2,000

1,253 / 2,000

Drebin Android apps Android app malware detectors

APP_C1APP_C2APP_C3

402102212

2,000 / 2,0002,000 / 2,0002,000 / 2,000

9/24/18

15

EmpiricalEvaluation

WhatHaveWeLearned?

• Feedingsameinputstodifferentimplementations

• Observingbehavioraldifferences

• Reportingdiscrepanciesasbugs

• Inputgenerationmethods– Unguidedvs. Guided– Tools:Csmith,NEZHA,DeepXPlore