PAGURUS: Low-Overhead Dynamic Information Flow ......PAGURUS: Low-Overhead Dynamic Information Flow Tracking on Loosely Coupled Accelerators Luca Piccolboni, Giuseppe Di Guglielmo

PAGURUS:Low-OverheadDynamicInformationFlowTrackingonLooselyCoupledAccelerators

LucaPiccolboni,GiuseppeDiGuglielmoandLucaP.CarloniColumbiaUniversity,NY,USA

ACM/IEEECODES+ISSS2018,Turin,Italy

ACM/IEEECODES+ISSS2018,Turin,Italy 2/16

[M.Gautschietal.,IEEEVLSI’17]

DataRAM

Instr.RAM

Boot.RAM

AXI

UART

SPIM.APB

PULPino

ProcessorCore(RI5CY)

Systems-on-Chip(SoCs)AreVulnerabletoSoftwareAttacks

ACM/IEEECODES+ISSS2018,Turin,Italy 3 /16

mainmemory

...

buff[0]=sw(7)

buff[1]=sw(7)

buff[9]=sw(7)

fun=0xAA

num=10

val=7

int buff[10], k;int (*fun)(int) = foo;int num = atoi(argv[1]);int val = atoi(argv[2]);/* this is a bad idea */for (k = 0; k < num; ++k)buff[k] = sw(val);

fun(1); // call foo?

Buffer-OverflowAttackAttackingPULPino

memorylocation:0xAA


mainmemory

...

buff[0]=sw(7)

buff[1]=sw(7)

buff[9]=sw(7)

fun=sw(7)

num=11

val=7


fun(1); // call foo?

memorylocation:0xAA

Buffer-OverflowAttackAttackingPULPino canbeusedtocalla

maliciousfunction


DynamicInformationFlowTracking(DIFT)

mainmemory

...

buff[0]=sw(0x7)

buff[1]=sw(0x7)

buff[9]=sw(0x7)

func=0xAA

num=0x7

val=0xA 1

0

1

1

1

1

mainmemory

...

buff[0]=sw(7)

buff[1]=sw(7)

buff[9]=sw(7)

num=7

val=11

AttackingPULPino


fun(1); // call fun

memorylocation:0xAA

fun=sw(7)

[G.E.Suhetal.,ACMASPLOS’04]tags



UART

SPIM.

PULPino

DataRAM

Instr.RAM

Boot.RAM


[C.Palmieroetal.,IEEEHPEC’18]DIFTExtensions

NowSecuredwithDIFTHomogenousSoCs

AXI

APB



UART

SPIM.

DataRAM

Instr.RAM

Boot.RAM


LooselyCoupledAccelerator#1

[C.Palmieroetal.,IEEEHPEC’18]

PULPinoNo-More-SecuredwithDIFTHeterogeneousSoCs

DIFTExtensions


AXI

APB


AttackingPULPino(Again)

int buff[10] = {0};int (*f)(int) = foo;int num = atoi(argv[1]);int val = atoi(argv[2]);/* this is a bad idea */hw(num, val, buff);

Buffer-OverflowAttack

mainmemory

...

buff[0]=sw(0x7)

buff[1]=sw(0x7)

buff[9]=sw(0x7)

func=0xAA

num=0x7

val=0xA 1

0

1

1

1

mainmemory

...

buff[0]=0

buff[1]=0

buff[9]=0

num=11

val=7

tags

fun=0xAA 1


AttackingPULPino(Again)

0

0

0

0

0theacceleratorisnotabletopropagatethetags

Buffer-OverflowAttack

1

mainmemory

...

buff[0]=sw(0x7)

buff[1]=sw(0x7)

buff[9]=sw(0x7)

func=0xAA

num=0x7

val=0xA

mainmemory

...

buff[0]=hw(7)

buff[1]=hw(7)

buff[9]=hw(7)

num=11

val=7

fun=hw(7)

canbeusedtocallamaliciousfunction

tags

int buff[10] = {0};int (*f)(int) = foo;int num = atoi(argv[1]);int val = atoi(argv[2]);/* this is a bad idea */hw(num, val, buff);


Contributions

1. WeproposePAGURUS,amethodologytodesignacircuitshell thataddsDIFTsupporttoaccelerators

DIFTShell


AXI

UART

SPIM.APB

PULPinoSystem-on-Chip

DataRAM

Instr.RAM

Boot.RAM


DIFTShe

ll

Contributions




Contributions

2. Weproposea metric toquantitativelymeasurethesecurityguaranteesprovidedbytheshell

a) Theshelldesignisindependent fromthedesignoftheacceleratorsandviceversa

b) Theshellhaslow overheadsonboththeperformanceandcostofaccelerators

1. WeproposePAGURUS,amethodologytodesignacircuitshell thataddsDIFTsupporttoaccelerators


1. Thehardwareissafe:nohardwareTrojans

AssumptionsandAttackModelPreliminaries

2. Thesoftwareisnot safe:itcontainsbugsandvulnerabilitiesusefulfortheattackers

TheattackersexploitthesevulnerabilitiesthroughcommonI/Ointerfaceswiththegoalofaffectingtheintegrityand/ortheconfidentialityofthehardware-acceleratedsoftwareapplications


mainmemory

value#1

1. CoupledScheme

value#2

value#3

PreliminariesTaggingScheme

tag#1

tag#2

tag#3

tags

[J.Porquetetal.,ACM/IEEECODES’13]


mainmemory

value#1

1. CoupledScheme

tag#1

tag#3

protectedregioninmemory

value#2

value#3

tag#2

2. DecoupledScheme




mainmemory

value#1

1. CoupledScheme

tag#1

tag#3

value#2

value#3

tag#2

2. DecoupledScheme2.1.InterleavedScheme

tagoffset =#wordsinmemorybetweentwoconsecutivevalues

(tagoffset=1)




Contributions



1. WeproposePAGURUS,amethodologytodesignacircuitshellthataddsDIFTsupporttoaccelerators


ArchitectureLooselyCoupledAccelerator

mainmemory

register#1

register#2

register#K

...

Accelerators

configuration reg#1 … reg#K

privatelocalmemory/scratchpad

bank bank bank bank


LooselyCoupledAccelerator

mainmemory

input

computeburstlength

Accelerators

configuration …

loadinputval val val



bank bank bank bank

Architecture



mainmemory

load computeloadinput

Accelerators

configuration

loadinputval val val val val val



bank bank bank bank

Architecture

output



mainmemory

load

storeburstlength storeoutput

Accelerators

load loadinputval val val val val val

computeloadinput

val val val



bank bank bank bank

Architecture


DIFTShellArchitecture

DIFTShe

ll

Accelerator




mainmemory

register#1

register#2

register#K

...

shellconfiguration

reg.#K+1:src_tag

reg.#K+2:dst_tag

dst_tagsrc_tag

Accelerator


DIFTShe

ll



shellconfiguration

mainmemory

input

src_tagsrc_tag shellloadlogic

if tag!=src_tagDIFT_exception!

val val

val tag val tag

Accelerator

dst_tagsrc_tagburstlength


DIFTShe

ll



shellconfiguration

mainmemory

shellloadlogic

shellstorelogic

val val

val tag val tag

output

dst_tagdst_tag

val tag val tag

dst_tagsrc_tag

Accelerator

burstlength


DIFTShe

ll


Contributions



1. WeproposePAGURUS,amethodologytodesignacircuitshellthataddsDIFTsupporttoaccelerators

2. Weproposea metric toquantitativelymeasurethesecurityguaranteesprovidedbytheshell


ASecurityMetricDefinition

mainmemory

input

value#1

src_tag

value#2

value#3


ASecurityMetricDefinition

mainmemory

input

value#1[overwritten]

src_tag[overwritten]



value#1


DIFTShe

ll

val tagval val

output


ASecurityMetric

mainmemory

input


src_tag[overwritten]




DIFTShe

ll

val tagval val

value#1

DIFT_exception!

InformationLeakage• Quantitativemetricforsecurity

Definition

output


ASecurityMetric

• InformationLeakage:amountofdatathatcanbeproducedasoutputbyanacceleratorbeforeitsshellrealizesthattheinputhasbeencorrupted

I/Oratio:thenumberofloadburstsnecessarytoproduceastoreburst

Analysis

1. Tagoffset: tagoffset leakage

2. Algorithm: I/Oratio leakage


ASecurityMetric

• InformationLeakage:amountofdatathatcanbeproducedasoutputbyanacceleratorbeforeitsshellrealizesthattheinputhasbeencorrupted

1. Tagoffset: tagoffset leakage

2. Algorithm: I/Oratio leakage

3. Implementation: burstlen. leakage

4. Workload: work. size leakage

Analysis

ExperimentalResults

13/16

ExperimentalSetup(1/2)


• Wedesignedthreelooselycoupledaccelerators:• GRAY:convertsaRGBimageintoagrayscaleimage• MEAN:calculatesthemeanofa2Dmatrix(columns)• MULTS:mutipliesa2Dmatrixbyitstranspose

ExperimentalResults

13/16




GRAY

loadburst

storeburst

ExperimentalResults

13/16




GRAY

…

MEAN

ExperimentalResults

13/16




GRAY

…

MEAN

…

MULTS

ExperimentalResults

13/16




• WeusedCadenceStratusHLS forhigh-levelsynthesisandXilinxVivadoforlogicsynthesisà Virtex-7FPGA

• WedesignedtheacceleratorsandtheshellinSystemC

ExperimentalResults

14/16



Weexploreddifferentalternativesbyvarying:• accelerator• tagoffset• burstsize• workload

[P.Mantovanietal.,ACM/IEEEDAC’16]

EmbeddedScalablePlatforms

[L.P.Carloni,ACM/IEEEDAC’16]


ProcessorCore (Leon3)

+Shell

MemoryController

I/Ochannelsandperipher.

Network-on-Chip

- 128x128- small- 512x512- medium- 2048x2048- large

ExperimentalResults

14/16ACM/IEEECODES+ISSS2018,Turin,Italy

QuantitativeSecurityAnalysis- MEAN

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rma

tion

lea

kag

e (

%)

burst size (bytes)

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

burst size (bytes)

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rmatio

n le

aka

ge (

%)

burst size (bytes)

medium

small

large

220

213

224

215216

211

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rma

tion

lea

kag

e (

%)

burst size (bytes)

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

burst size (bytes)

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rmatio

n le

aka

ge (

%)

burst size (bytes)

ExperimentalResults



maxinformationleakage=>thehighesttagoffset

mininformationleakage=>thelowesttagoffset

220

213

medium

224

215

small

large

216

211

ExperimentalResults



0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rma

tion

lea

kag

e (

%)

burst size (bytes)

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

burst size (bytes)

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rmatio

n le

aka

ge (

%)

burst size (bytes)

220

213

medium

224

215

small

large

216

211

ExperimentalResults


QuantitativeSecurityAnalysis- GRAY

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rmatio

n le

aka

ge (

%)

burst size (bytes)

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rmatio

n le

aka

ge (

%)

burst size (bytes)

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rma

tion

lea

kag

e (

%)

burst size (bytes)

220

213

medium

224

215

small

large

216

211

ExperimentalResults


QuantitativeSecurityAnalysis- GRAY

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rma

tion

lea

kag

e (

%)

burst size (bytes)

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rmatio

n le

aka

ge (

%)

burst size (bytes)

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rmatio

n le

aka

ge (

%)

burst size (bytes)

219

25

medium

223

25

small

large

215

25

ExperimentalResults


QuantitativeSecurityAnalysis- MULTS

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

burst size (bytes)

0.00%

0.02%

0.04%

0.06%

0.08%

0.10%

0.12%

0.14%

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

burst size (bytes)

0.0%

0.1%

0.2%

0.3%

0.4%

0.5%

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rmatio

n le

aka

ge (

%)

burst size (bytes)

0.0%

0.2%

0.4%

0.6%

0.8%

1.0%

1.2%

1.4%

1.6%

220

214

medium

224

216

small

large

216

212

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

burst size (bytes)

0.00%

0.02%

0.04%

0.06%

0.08%

0.10%

0.12%

0.14%

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

info

rmatio

n le

aka

ge (

%)

burst size (bytes)

0.0%

0.2%

0.4%

0.6%

0.8%

1.0%

1.2%

1.4%

1.6%

0%

20%

40%

60%

80%

100%

26 27 28 29 210 211 212 213

burst size (bytes)

0.0%

0.1%

0.2%

0.3%

0.4%

0.5%

ExperimentalResults


QuantitativeSecurityAnalysis- MULTS

220

214

medium

224

216

small

large

216

212

ExperimentalResults


PerformanceAnalysis- GRAY

1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

2.6

26 27 28 29 210 211 212 213

gray - large

no

rma

lize

d e

xecu

tion

tim

e

burst size (bytes)

1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

26 27 28 29 210 211 212 213

gray - medium

norm

aliz

ed e

xecu

tion tim

e

burst size (bytes)

1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

26 27 28 29 210 211 212 213

gray - small

norm

aliz

ed e

xecu

tion tim

e

burst size (bytes)

20

26

212

no tags

medium

small

large

ExperimentalResults


1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

2.6

26 27 28 29 210 211 212 213

gray - large

no

rma

lize

d e

xecu

tion

tim

e

burst size (bytes)

1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

26 27 28 29 210 211 212 213

gray - medium

norm

aliz

ed e

xecu

tion tim

e

burst size (bytes)

1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

26 27 28 29 210 211 212 213

gray - small

norm

aliz

ed e

xecu

tion tim

e

burst size (bytes)

20

26

212

no tagsPerformanceAnalysis- GRAYmedium

small

large

ExperimentalResults


1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

2.6

26 27 28 29 210 211 212 213

gray - large

no

rma

lize

d e

xecu

tion

tim

e

burst size (bytes)

1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

26 27 28 29 210 211 212 213

gray - medium

norm

aliz

ed e

xecu

tion tim

e

burst size (bytes)

1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

26 27 28 29 210 211 212 213

gray - small

norm

aliz

ed e

xecu

tion tim

e

burst size (bytes)

20

26

212


small

large

ExperimentalResults


1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

2.6

26 27 28 29 210 211 212 213

gray - large

no

rma

lize

d e

xecu

tion

tim

e

burst size (bytes)

1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

26 27 28 29 210 211 212 213

gray - small

norm

aliz

ed e

xecu

tion tim

e

burst size (bytes)

1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

26 27 28 29 210 211 212 213

gray - medium

norm

aliz

ed e

xecu

tion tim

e

burst size (bytes)

20

26

212


small

large

15/16

• WeproposePAGURUS,aflexiblemethodologytodesignashellthatextendsDIFTtoaccelerators

1. Theshelldesignisindependentfromtheacceleratordesignandviceversa

2. Theshellhasnegligiblecostoverheadandreasonableperformanceoverhead

• Wedefinethemetric ofinformationleakageforacceleratorstoquantitativelymeasuresecurity


Conclusions

Speaker:LucaPiccolboniColumbiaUniversity,NY

Questions?


PAGURUS:Low-OverheadDynamicInformationFlowTrackingonLooselyCoupleAccelerators

PAGURUS: Low-Overhead Dynamic Information Flow ......PAGURUS: Low-Overhead Dynamic Information Flow Tracking on Loosely Coupled Accelerators Luca Piccolboni, Giuseppe Di Guglielmo

Documents