security for machine learning systems · computer architectureand robust energy-efficient...
TRANSCRIPT
-
Computer Architecture and RobustEnergy-Efficient Technologies Group
CAREech
Security for Machine Learning Systems Attacks, Defenses, & Reconfigurability
M. Shafique
-
2
Who Ruled the World!2
Age of PowerMan-Power (#), Skills, Strength, Courage, etc.
Age of Resources and IndustryFuel, Industrial Tech., Economic Politics, etc.
Age of Data and AIData is the New Fuel
Innovation in Technology is the New PoliticsNation-wide Race for Dominance in AI
-
3
Smart Cyber Physical Systems & Internet-of-Things
Smart Traffic ControlSmart Transport Systems
Smart Gridshttp://solutions.3m.com/wps/portal/3M/en_EU/Sma
rtGrid/EU-Smart-Grid/
https://www.automotiveworld.com/analysis/automotive-cyber-physical-systems-next-computing-revolution/
https://www.emaze.com/@ACIOWOWR/IMSA-Slide-Show
Industry 4.0: Smart Industrial Automation
https://vimeo.com/145877805
Smart Houseshttps://www.linkedin.com/pulse/smart-homes-private-secure-future-intelligent-home-tripti-jha
Smart Automobileshttp://www.it5g.com/latest-software-enhancements-in-the-auto-industry/
Smart Robotshttp://alpha-smart.com/alphaboten
Smart Health Care
Applications
http://solutions.3m.com/wps/portal/3M/en_EU/SmartGrid/EU-Smart-Grid/https://www.automotiveworld.com/analysis/automotive-cyber-physical-systems-next-computing-revolution/https://www.emaze.com/@ACIOWOWR/IMSA-Slide-Showhttps://vimeo.com/145877805https://www.linkedin.com/pulse/smart-homes-private-secure-future-intelligent-home-tripti-jhahttp://www.it5g.com/latest-software-enhancements-in-the-auto-industry/http://alpha-smart.com/alphaboten
-
4
Smart CPS & IoT => The Big Data Processing Challenge!4
By 2025: 75 Billion connected devices By 2025: 4,785 interactions per
connected person in one day.
By 2025: 160 Zettabyte measured data By 2020: for every person on earth,
1.7 MB data/sec will be generated.
-
5
Smart CPS & IoT => The Big Data Processing Challenge!5
By 2025: 75 Billion connected devices By 2025: 4,785 interactions per
connected person in one day.
By 2025: 160 Zettabyte measured data By 2020: for every person on earth,
1.7 MB data/sec will be generated.
AI / ML is inevitable, we have to efficiently infer knowledge from the
big data, and derive predictions
-
6
ML Applications => requiring High Efficiency Gains6
Autonomous Driving
Cancer DetectionNatural Language
Processing
Image Classification
Object Detection & Localization
Machine Translation
Strategy Games
Forex/Stocks Trading
-
7
Machine Learning and the Computing Efficiency Gap!
Deep Blue vs. Garry Kasparov AlphaGO vs. Lee Sedol
IBM RS/6000 SP high-performance computer
1996 2016
IBM Watson vs. Brad Rutter and Ken Jennings
IBM Blue Gene supercomputer
1,920 CPUs and 280 GPUs
2011
Google's cloud computing
2,880 POWER7 processor threads and 16 terabytes of
RAM
Data Source: https://jacquesmattheij.com/another-way-of-looking-at-lee-sedol-vs-alphago
30 P2SC nodes + 480
dedicated ASICs to play
chess
1MW
200KW
20W900W + Power for 480 dedicated ASICs
-
8
Machine Learning and the Computing Efficiency Gap!
Deep Blue vs. Garry Kasparov AlphaGO vs. Lee Sedol
IBM RS/6000 SP high-performance computer
1996 2016
IBM Watson vs. Brad Rutter and Ken Jennings
IBM Blue Gene supercomputer
1,920 CPUs and 280 GPUs
2011
Google's cloud computing
2,880 POWER7 processor threads and 16 terabytes of
RAM
Data Source: https://jacquesmattheij.com/another-way-of-looking-at-lee-sedol-vs-alphago
30 P2SC nodes + 480
dedicated ASICs to play
chess
1MW
200KW
20W900W + Power for 480 dedicated ASICs
We need 10,000x Gains to Bridge this Gap!!!
-
9
Brain-Inspired Computing: [email protected]
Source: S. Mitra, P. Wong (Stanford), C. Mackin (MIT), J. Zhang (Google)
Software (Multi-Cores, GPUs)
Approximate Computing
Deep Learning Architectures
Neuromorphic Architectures and Beyond
Post-CMOS Technologies
......
... ... ... ...
Accurate SAD Accelerators
...
...
... ... ... ...
Approximate SAD Accel. Variant1
... ......
...
... ... ... ...
Approximate SAD Accel. VariantN-1
...
...
... ... ... ...
Approximate SAD Accel. VariantN
SAD Accelerator Array Output and Monitoring
(MV, SAD, etc.)
AGU
On-Chip Memory
Ref. FrameSearch Window
Current Frame CTU
Power-Gating Control Approximate Variant Selection Unit
User ConstraintsCPU Main Memory
System Bus
CPU executes the ME
algorithm andHEVC
Memory stores the video frames
1 uJ/neuron
FC layer
C1
1234567890
C2P1 P2 FC
Input Image
First ConvolutionC1@6 filters
PoolingP1@Average
Second ConvolutionC2@12 filters
PoolingP2@Average
Source: Huawei Kirin Chip
Source: IBM, TrueNorth Chip
??
In-Memory Computing
Source: IBM Research
CPS
High Performance, Energy Efficiency,
Configurability, Reliability & Security
*
...
...
.
.
.
.
.
.
.
.
.
.
.
.
Accurate SAD
Accelerators
...
...
.
.
.
.
.
.
.
.
.
.
.
.
Approximate SAD
Accel. Variant
1
.
.
.
.
.
.
...
...
.
.
.
.
.
.
.
.
.
.
.
.
Approximate SAD
Accel. Variant
N-1
...
...
.
.
.
.
.
.
.
.
.
.
.
.
Approximate SAD
Accel. Variant
N
SAD Accelerator Array
Output and
Monitoring
(MV, SAD, etc.)
AGU
On-Chip Memory
Ref.
Frame
Search
Window
Current
Frame
CTU
Power-Gating Control
Approximate Variant Selection Unit
User
Constraints
CPUMain Memory
System Bus
CPU
executes
the ME
algorithm
andHEVC
Memory stores
the video frames
-
10
Robustness for Machine Learning: News Feed
Hackers trick a Tesla into veering into the wrong lane
https://www.technologyreview.com/f/613254/hackers-trick-teslas-autopilot-into-
veering-towards-oncoming-traffic/
https://www.youtube.com/watch?v=a7L51u23YoM
Self-driving car crash in Arizona: Waymo van involved in Chandler
collision
https://www.technologyreview.com/f/613254/hackers-trick-teslas-autopilot-into-veering-towards-oncoming-traffic/https://www.youtube.com/watch?v=a7L51u23YoM
-
11
Risks that Arise from Adversarial Examples
FACE RECOGNITION“Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition” [Sharif et. al. CCS 2016]
SPEECH RECOGNITION“Hidden Voice Commands”, [Carlini et. al. USENIX 2016]
MALWARE DETECTION“Automatically Evading Classifiers: A Case Study on PDF Malware Classifiers” [Xu et. al. NDSS2016 ]
TEXT CLASSIFICATION (Spam)“Crafting Adversarial Input Sequences for Recurrent Neural Networks” [Papernot et. al. 2016]
Junaid, Shafique et al. VTC 2019 Tutoria on Adv. ML
-
12
Reliability and Security for Machine Learning Systems
DrainSource
p+ p+
n – substrate
GateOxide Layer
Vg= – Vdd
Si HTRAP
OH+
NBTI
Aging
HCID
Process Variations Soft Errors
n+ n+
P-Well
P-Substrate
IsolationGate
+-+-
+-+- +-
+- +- +-
+-
+- Depletion Region
High-Energy Particle (Neutron or Proton)
Side Channel Attacks
1 0 1 1 0
ProcessingComputations
Mem
oryPower Supply
Machine Learning-based System Src: google ImageHardware Trojans
Neural Trojans
• M. A. Hanif, F. Khalid, R. V. W. Putra, S. Rehman, M. Shafique, “Robust Machine Learning Systems: Reliability and Security for Deep Neural Networks”, in IOLTS-2018, Platja d'Aro, Spain, pp. 257 - 260.
• F. Kriebel, S. Rehman, M. A. Hanif, F. Khalid, M. Shafique, “Robustness for Smart Cyber-Physical Systems and Internet-of-Things: From Adaptive Robustness Methods to Reliability and Security for Machine Learning”, ISVLSI-2018, Hong Kong, China, pp. 581-586.
Adversarial Attacks
Stop Sign Speed Limit (60km/h)
+
Backdoor Attacks
Src: Reiter@CCS2016
Model Stealing Attacks
Src: Tramer@UsenixSec16
P-type MOSFET
DrainSourcep+p+n – substrateGateOxide LayerVg= – VddSiHTRAPOH+
Drain
Graph3
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.5074
4.50062
4.51561
4.50764
4.51668
4.50609
4.50268
4.52268
4.50481
4.51454
4.52499
4.51218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.51268
Sheet1
160192224256
14.49744.497400
24.500624.5006200
34.515614.5156100
44.497644.4976400
54.512684.5126851
64.506094.50609279
74.512684.512687152
84.512684.51268171111
94.494814.49481271215
104.514544.51454361324
114.524994.52499507471
124.502184.50218594580
134.530264.53026656627
144.517564.51756753676
154.506534.50653788775
164.502184.50218850821
174.485244.48524878841
184.517564.51756921902
194.494814.49481931918
204.4914.491957924
214.4934.493977950
224.491884.49188973950
234.49744.4974973967
244.500624.50062976972
254.515614.51561988975
264.497644.49764988975
274.512684.5074989980
284.506094.50062997985
294.512684.51561993983
304.512684.50764989985
314.494814.51668994995
324.514544.50609991989
334.524994.50268997991
344.502184.52268993992
354.530264.50481999997
364.517564.51454995994
374.506534.52499994996
384.502184.51218995991
394.485244.530261000997
404.517564.51756995992
414.494814.50653998990
424.4914.50218997991
434.4934.48524997996
444.491884.51756999990
454.49744.49481997997
464.500624.491999997
474.515614.493999997
484.497644.49188999997
494.512684.51268999995
504.50609999997
514.51268997998
524.512681000996
534.49481997994
544.514541000992
554.52499994998
564.5021810001000
574.53026998988
584.51756997996
594.50653999996
604.50218996999
614.48524999996
624.517561000994
634.49481997998
644.491999999
654.493998994
664.49188992994
674.4974998998
684.50062996999
694.515619991000
704.49764998996
714.51268996997
724.50609994994
734.51268999995
744.5126810001000
754.49481999994
764.51454999997
774.52499998998
784.50218985998
794.53026997982
804.517569971000
814.506539991000
824.50218998997
834.4852410001000
844.517569991000
854.494811000998
864.4919991000
874.4931000990
884.49188999996
894.4974999994
904.500621000998
914.515611000999
924.49764997999
934.512681000995
944.506099981000
954.51268998997
964.512681000999
974.49481999997
984.51454998996
994.52499997997
1004.502189961000
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
Sheet1
160
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.51268
Sheet2
192
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.5074
4.50062
4.51561
4.50764
4.51668
4.50609
4.50268
4.52268
4.50481
4.51454
4.52499
4.51218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.51268
Sheet3
160192224256160192
14.49744.4974004.49744.4974
24.500624.50062004.500624.50062
34.515614.51561004.515614.51561
44.497644.49764004.497644.49764
54.512684.51268514.512684.51268
64.506094.506092794.506094.50609
74.512684.5126871524.512684.51268
84.512684.51268171111double+add4.512684.51268
94.494814.494812712154.494814.49481
104.514544.514543613244.514544.51454
114.524994.524995074714.524994.52499
124.502184.502185945804.502184.50218
134.530264.530266566274.530264.53026
144.517564.517567536764.517564.51756
154.506534.506537887754.506534.50653
164.502184.502188508214.502184.50218
174.485244.485248788414.485244.48524
184.517564.517569219024.517564.51756
194.494814.494819319184.494814.49481
204.4914.4919579244.4914.491
214.4934.4939779504.4934.493
224.491884.491889739504.49744.49188
234.49744.4974973967double+double4.500624.4974
244.500624.500629769724.515614.50062
254.515614.515619889754.497644.51561
264.497644.497649889754.512684.49764
274.512684.50749899804.506094.5074
284.506094.500629979854.512684.50062
294.512684.515619939834.512684.51561
304.512684.507649899854.494814.50764
314.494814.516689949954.514544.51668
324.514544.506099919894.524994.50609
334.524994.502689979914.502184.50268
344.502184.522689939924.530264.52268
354.530264.504819999974.517564.50481
364.517564.514549959944.506534.51454
374.506534.524999949964.502184.52499
384.502184.512189959914.485244.51218
394.485244.5302610009974.517564.53026
404.517564.517569959924.494814.51756
414.494814.506539989904.4914.50653
424.4914.502189979914.4934.50218
434.4934.48524997996tau+add4.48524
444.491884.517569999904.51756
454.49744.494819979974.49481
464.500624.4919999974.491
474.515614.4939999974.493
484.497644.491889999974.49188
494.512684.512689999954.51268
504.50609999997
514.51268997998
524.512681000996
534.49481997994
544.514541000992
554.52499994998
564.5021810001000
574.53026998988
584.51756997996
594.50653999996
604.50218996999
614.48524999996
624.517561000994tau+tau
634.49481997998
644.491999999
654.493998994
664.49188992994
674.4974998998
684.50062996999
694.5156199910004.50218
704.497649989964.48524
714.512689969974.51756
724.506099949944.49481
734.512689999954.50218
744.51268100010004.48524
754.494819999944.51756
764.514549999974.49481
774.52499998998
784.50218985998
794.53026997982
804.517569971000
814.506539991000
824.502189989974.50218
834.48524100010004.48524
844.5175699910004.51756
854.4948110009984.49481
864.49199910004.493
874.49310009904.49188
884.491889999964.4974
894.49749999944.50062
904.5006210009984.51561
914.5156110009994.49764
924.497649979994.5074
934.5126810009954.50062
944.5060999810004.51561
954.512689989974.50764
964.5126810009994.51668
974.494819999974.50609
984.514549989964.50268
994.524999979974.52268
1004.5021899610004.50481
4.530264.51454
4.517564.52499
4.506534.51218
4.502184.53026
4.485244.51756
4.517564.50653
4.494814.50218
4.4914.48524
4.4934.51756
4.491884.49481
4.491
4.493
4.49188
4.51268
Sheet3
Graph4
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
Sheet1
160192224256
14.49744.497400
24.500624.5006200
34.515614.5156100
44.497644.4976400
54.512684.5126851
64.506094.50609279
74.512684.512687152
84.512684.51268171111
94.494814.49481271215
104.514544.51454361324
114.524994.52499507471
124.502184.50218594580
134.530264.53026656627
144.517564.51756753676
154.506534.50653788775
164.502184.50218850821
174.485244.48524878841
184.517564.51756921902
194.494814.49481931918
204.4914.491957924
214.4934.493977950
224.491884.49188973950
234.49744.4974973967
244.500624.50062976972
254.515614.51561988975
264.497644.49764988975
274.512684.5074989980
284.506094.50062997985
294.512684.51561993983
304.512684.50764989985
314.494814.51668994995
324.514544.50609991989
334.524994.50268997991
344.502184.52268993992
354.530264.50481999997
364.517564.51454995994
374.506534.52499994996
384.502184.51218995991
394.485244.530261000997
404.517564.51756995992
414.494814.50653998990
424.4914.50218997991
434.4934.48524997996
444.491884.51756999990
454.49744.49481997997
464.500624.491999997
474.515614.493999997
484.497644.49188999997
494.512684.51268999995
504.50609999997
514.51268997998
524.512681000996
534.49481997994
544.514541000992
554.52499994998
564.5021810001000
574.53026998988
584.51756997996
594.50653999996
604.50218996999
614.48524999996
624.517561000994
634.49481997998
644.491999999
654.493998994
664.49188992994
674.4974998998
684.50062996999
694.515619991000
704.49764998996
714.51268996997
724.50609994994
734.51268999995
744.5126810001000
754.49481999994
764.51454999997
774.52499998998
784.50218985998
794.53026997982
804.517569971000
814.506539991000
824.50218998997
834.4852410001000
844.517569991000
854.494811000998
864.4919991000
874.4931000990
884.49188999996
894.4974999994
904.500621000998
914.515611000999
924.49764997999
934.512681000995
944.506099981000
954.51268998997
964.512681000999
974.49481999997
984.51454998996
994.52499997997
1004.502189961000
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
Sheet1
160
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.51268
Sheet2
192
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.5074
4.50062
4.51561
4.50764
4.51668
4.50609
4.50268
4.52268
4.50481
4.51454
4.52499
4.51218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.51268
Sheet3
160192224256160192
14.49744.4974004.49744.4974
24.500624.50062004.500624.50062
34.515614.51561004.515614.51561
44.497644.49764004.497644.49764
54.512684.51268514.512684.51268
64.506094.506092794.506094.50609
74.512684.5126871524.512684.51268
84.512684.51268171111double+add4.512684.51268
94.494814.494812712154.494814.49481
104.514544.514543613244.514544.51454
114.524994.524995074714.524994.52499
124.502184.502185945804.502184.50218
134.530264.530266566274.530264.53026
144.517564.517567536764.517564.51756
154.506534.506537887754.506534.50653
164.502184.502188508214.502184.50218
174.485244.485248788414.485244.48524
184.517564.517569219024.517564.51756
194.494814.494819319184.494814.49481
204.4914.4919579244.4914.491
214.4934.4939779504.4934.493
224.491884.491889739504.49744.49188
234.49744.4974973967double+double4.500624.4974
244.500624.500629769724.515614.50062
254.515614.515619889754.497644.51561
264.497644.497649889754.512684.49764
274.512684.50749899804.506094.5074
284.506094.500629979854.512684.50062
294.512684.515619939834.512684.51561
304.512684.507649899854.494814.50764
314.494814.516689949954.514544.51668
324.514544.506099919894.524994.50609
334.524994.502689979914.502184.50268
344.502184.522689939924.530264.52268
354.530264.504819999974.517564.50481
364.517564.514549959944.506534.51454
374.506534.524999949964.502184.52499
384.502184.512189959914.485244.51218
394.485244.5302610009974.517564.53026
404.517564.517569959924.494814.51756
414.494814.506539989904.4914.50653
424.4914.502189979914.4934.50218
434.4934.48524997996tau+add4.48524
444.491884.517569999904.51756
454.49744.494819979974.49481
464.500624.4919999974.491
474.515614.4939999974.493
484.497644.491889999974.49188
494.512684.512689999954.51268
504.50609999997
514.51268997998
524.512681000996
534.49481997994
544.514541000992
554.52499994998
564.5021810001000
574.53026998988
584.51756997996
594.50653999996
604.50218996999
614.48524999996
624.517561000994tau+tau
634.49481997998
644.491999999
654.493998994
664.49188992994
674.4974998998
684.50062996999
694.5156199910004.50218
704.497649989964.48524
714.512689969974.51756
724.506099949944.49481
734.512689999954.50218
744.51268100010004.48524
754.494819999944.51756
764.514549999974.49481
774.52499998998
784.50218985998
794.53026997982
804.517569971000
814.506539991000
824.502189989974.50218
834.48524100010004.48524
844.5175699910004.51756
854.4948110009984.49481
864.49199910004.493
874.49310009904.49188
884.491889999964.4974
894.49749999944.50062
904.5006210009984.51561
914.5156110009994.49764
924.497649979994.5074
934.5126810009954.50062
944.5060999810004.51561
954.512689989974.50764
964.5126810009994.51668
974.494819999974.50609
984.514549989964.50268
994.524999979974.52268
1004.5021899610004.50481
4.530264.51454
4.517564.52499
4.506534.51218
4.502184.53026
4.485244.51756
4.517564.50653
4.494814.50218
4.4914.48524
4.4934.51756
4.491884.49481
4.491
4.493
4.49188
4.51268
Sheet3
Graph3
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.5074
4.50062
4.51561
4.50764
4.51668
4.50609
4.50268
4.52268
4.50481
4.51454
4.52499
4.51218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.51268
Sheet1
160192224256
14.49744.497400
24.500624.5006200
34.515614.5156100
44.497644.4976400
54.512684.5126851
64.506094.50609279
74.512684.512687152
84.512684.51268171111
94.494814.49481271215
104.514544.51454361324
114.524994.52499507471
124.502184.50218594580
134.530264.53026656627
144.517564.51756753676
154.506534.50653788775
164.502184.50218850821
174.485244.48524878841
184.517564.51756921902
194.494814.49481931918
204.4914.491957924
214.4934.493977950
224.491884.49188973950
234.49744.4974973967
244.500624.50062976972
254.515614.51561988975
264.497644.49764988975
274.512684.5074989980
284.506094.50062997985
294.512684.51561993983
304.512684.50764989985
314.494814.51668994995
324.514544.50609991989
334.524994.50268997991
344.502184.52268993992
354.530264.50481999997
364.517564.51454995994
374.506534.52499994996
384.502184.51218995991
394.485244.530261000997
404.517564.51756995992
414.494814.50653998990
424.4914.50218997991
434.4934.48524997996
444.491884.51756999990
454.49744.49481997997
464.500624.491999997
474.515614.493999997
484.497644.49188999997
494.512684.51268999995
504.50609999997
514.51268997998
524.512681000996
534.49481997994
544.514541000992
554.52499994998
564.5021810001000
574.53026998988
584.51756997996
594.50653999996
604.50218996999
614.48524999996
624.517561000994
634.49481997998
644.491999999
654.493998994
664.49188992994
674.4974998998
684.50062996999
694.515619991000
704.49764998996
714.51268996997
724.50609994994
734.51268999995
744.5126810001000
754.49481999994
764.51454999997
774.52499998998
784.50218985998
794.53026997982
804.517569971000
814.506539991000
824.50218998997
834.4852410001000
844.517569991000
854.494811000998
864.4919991000
874.4931000990
884.49188999996
894.4974999994
904.500621000998
914.515611000999
924.49764997999
934.512681000995
944.506099981000
954.51268998997
964.512681000999
974.49481999997
984.51454998996
994.52499997997
1004.502189961000
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
Sheet1
160
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.51268
Sheet2
192
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.5074
4.50062
4.51561
4.50764
4.51668
4.50609
4.50268
4.52268
4.50481
4.51454
4.52499
4.51218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.51268
Sheet3
160192224256160192
14.49744.4974004.49744.4974
24.500624.50062004.500624.50062
34.515614.51561004.515614.51561
44.497644.49764004.497644.49764
54.512684.51268514.512684.51268
64.506094.506092794.506094.50609
74.512684.5126871524.512684.51268
84.512684.51268171111double+add4.512684.51268
94.494814.494812712154.494814.49481
104.514544.514543613244.514544.51454
114.524994.524995074714.524994.52499
124.502184.502185945804.502184.50218
134.530264.530266566274.530264.53026
144.517564.517567536764.517564.51756
154.506534.506537887754.506534.50653
164.502184.502188508214.502184.50218
174.485244.485248788414.485244.48524
184.517564.517569219024.517564.51756
194.494814.494819319184.494814.49481
204.4914.4919579244.4914.491
214.4934.4939779504.4934.493
224.491884.491889739504.49744.49188
234.49744.4974973967double+double4.500624.4974
244.500624.500629769724.515614.50062
254.515614.515619889754.497644.51561
264.497644.497649889754.512684.49764
274.512684.50749899804.506094.5074
284.506094.500629979854.512684.50062
294.512684.515619939834.512684.51561
304.512684.507649899854.494814.50764
314.494814.516689949954.514544.51668
324.514544.506099919894.524994.50609
334.524994.502689979914.502184.50268
344.502184.522689939924.530264.52268
354.530264.504819999974.517564.50481
364.517564.514549959944.506534.51454
374.506534.524999949964.502184.52499
384.502184.512189959914.485244.51218
394.485244.5302610009974.517564.53026
404.517564.517569959924.494814.51756
414.494814.506539989904.4914.50653
424.4914.502189979914.4934.50218
434.4934.48524997996tau+add4.48524
444.491884.517569999904.51756
454.49744.494819979974.49481
464.500624.4919999974.491
474.515614.4939999974.493
484.497644.491889999974.49188
494.512684.512689999954.51268
504.50609999997
514.51268997998
524.512681000996
534.49481997994
544.514541000992
554.52499994998
564.5021810001000
574.53026998988
584.51756997996
594.50653999996
604.50218996999
614.48524999996
624.517561000994tau+tau
634.49481997998
644.491999999
654.493998994
664.49188992994
674.4974998998
684.50062996999
694.5156199910004.50218
704.497649989964.48524
714.512689969974.51756
724.506099949944.49481
734.512689999954.50218
744.51268100010004.48524
754.494819999944.51756
764.514549999974.49481
774.52499998998
784.50218985998
794.53026997982
804.517569971000
814.506539991000
824.502189989974.50218
834.48524100010004.48524
844.5175699910004.51756
854.4948110009984.49481
864.49199910004.493
874.49310009904.49188
884.491889999964.4974
894.49749999944.50062
904.5006210009984.51561
914.5156110009994.49764
924.497649979994.5074
934.5126810009954.50062
944.5060999810004.51561
954.512689989974.50764
964.5126810009994.51668
974.494819999974.50609
984.514549989964.50268
994.524999979974.52268
1004.5021899610004.50481
4.530264.51454
4.517564.52499
4.506534.51218
4.502184.53026
4.485244.51756
4.517564.50653
4.494814.50218
4.4914.48524
4.4934.51756
4.491884.49481
4.491
4.493
4.49188
4.51268
Sheet3
Graph3
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.5074
4.50062
4.51561
4.50764
4.51668
4.50609
4.50268
4.52268
4.50481
4.51454
4.52499
4.51218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.51268
Sheet1
160192224256
14.49744.497400
24.500624.5006200
34.515614.5156100
44.497644.4976400
54.512684.5126851
64.506094.50609279
74.512684.512687152
84.512684.51268171111
94.494814.49481271215
104.514544.51454361324
114.524994.52499507471
124.502184.50218594580
134.530264.53026656627
144.517564.51756753676
154.506534.50653788775
164.502184.50218850821
174.485244.48524878841
184.517564.51756921902
194.494814.49481931918
204.4914.491957924
214.4934.493977950
224.491884.49188973950
234.49744.4974973967
244.500624.50062976972
254.515614.51561988975
264.497644.49764988975
274.512684.5074989980
284.506094.50062997985
294.512684.51561993983
304.512684.50764989985
314.494814.51668994995
324.514544.50609991989
334.524994.50268997991
344.502184.52268993992
354.530264.50481999997
364.517564.51454995994
374.506534.52499994996
384.502184.51218995991
394.485244.530261000997
404.517564.51756995992
414.494814.50653998990
424.4914.50218997991
434.4934.48524997996
444.491884.51756999990
454.49744.49481997997
464.500624.491999997
474.515614.493999997
484.497644.49188999997
494.512684.51268999995
504.50609999997
514.51268997998
524.512681000996
534.49481997994
544.514541000992
554.52499994998
564.5021810001000
574.53026998988
584.51756997996
594.50653999996
604.50218996999
614.48524999996
624.517561000994
634.49481997998
644.491999999
654.493998994
664.49188992994
674.4974998998
684.50062996999
694.515619991000
704.49764998996
714.51268996997
724.50609994994
734.51268999995
744.5126810001000
754.49481999994
764.51454999997
774.52499998998
784.50218985998
794.53026997982
804.517569971000
814.506539991000
824.50218998997
834.4852410001000
844.517569991000
854.494811000998
864.4919991000
874.4931000990
884.49188999996
894.4974999994
904.500621000998
914.515611000999
924.49764997999
934.512681000995
944.506099981000
954.51268998997
964.512681000999
974.49481999997
984.51454998996
994.52499997997
1004.502189961000
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
Sheet1
160
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.51268
Sheet2
192
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.5074
4.50062
4.51561
4.50764
4.51668
4.50609
4.50268
4.52268
4.50481
4.51454
4.52499
4.51218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.51268
Sheet3
160192224256160192
14.49744.4974004.49744.4974
24.500624.50062004.500624.50062
34.515614.51561004.515614.51561
44.497644.49764004.497644.49764
54.512684.51268514.512684.51268
64.506094.506092794.506094.50609
74.512684.5126871524.512684.51268
84.512684.51268171111double+add4.512684.51268
94.494814.494812712154.494814.49481
104.514544.514543613244.514544.51454
114.524994.524995074714.524994.52499
124.502184.502185945804.502184.50218
134.530264.530266566274.530264.53026
144.517564.517567536764.517564.51756
154.506534.506537887754.506534.50653
164.502184.502188508214.502184.50218
174.485244.485248788414.485244.48524
184.517564.517569219024.517564.51756
194.494814.494819319184.494814.49481
204.4914.4919579244.4914.491
214.4934.4939779504.4934.493
224.491884.491889739504.49744.49188
234.49744.4974973967double+double4.500624.4974
244.500624.500629769724.515614.50062
254.515614.515619889754.497644.51561
264.497644.497649889754.512684.49764
274.512684.50749899804.506094.5074
284.506094.500629979854.512684.50062
294.512684.515619939834.512684.51561
304.512684.507649899854.494814.50764
314.494814.516689949954.514544.51668
324.514544.506099919894.524994.50609
334.524994.502689979914.502184.50268
344.502184.522689939924.530264.52268
354.530264.504819999974.517564.50481
364.517564.514549959944.506534.51454
374.506534.524999949964.502184.52499
384.502184.512189959914.485244.51218
394.485244.5302610009974.517564.53026
404.517564.517569959924.494814.51756
414.494814.506539989904.4914.50653
424.4914.502189979914.4934.50218
434.4934.48524997996tau+add4.48524
444.491884.517569999904.51756
454.49744.494819979974.49481
464.500624.4919999974.491
474.515614.4939999974.493
484.497644.491889999974.49188
494.512684.512689999954.51268
504.50609999997
514.51268997998
524.512681000996
534.49481997994
544.514541000992
554.52499994998
564.5021810001000
574.53026998988
584.51756997996
594.50653999996
604.50218996999
614.48524999996
624.517561000994tau+tau
634.49481997998
644.491999999
654.493998994
664.49188992994
674.4974998998
684.50062996999
694.5156199910004.50218
704.497649989964.48524
714.512689969974.51756
724.506099949944.49481
734.512689999954.50218
744.51268100010004.48524
754.494819999944.51756
764.514549999974.49481
774.52499998998
784.50218985998
794.53026997982
804.517569971000
814.506539991000
824.502189989974.50218
834.48524100010004.48524
844.5175699910004.51756
854.4948110009984.49481
864.49199910004.493
874.49310009904.49188
884.491889999964.4974
894.49749999944.50062
904.5006210009984.51561
914.5156110009994.49764
924.497649979994.5074
934.5126810009954.50062
944.5060999810004.51561
954.512689989974.50764
964.5126810009994.51668
974.494819999974.50609
984.514549989964.50268
994.524999979974.52268
1004.5021899610004.50481
4.530264.51454
4.517564.52499
4.506534.51218
4.502184.53026
4.485244.51756
4.517564.50653
4.494814.50218
4.4914.48524
4.4934.51756
4.491884.49481
4.491
4.493
4.49188
4.51268
Sheet3
Graph4
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
Sheet1
160192224256
14.49744.497400
24.500624.5006200
34.515614.5156100
44.497644.4976400
54.512684.5126851
64.506094.50609279
74.512684.512687152
84.512684.51268171111
94.494814.49481271215
104.514544.51454361324
114.524994.52499507471
124.502184.50218594580
134.530264.53026656627
144.517564.51756753676
154.506534.50653788775
164.502184.50218850821
174.485244.48524878841
184.517564.51756921902
194.494814.49481931918
204.4914.491957924
214.4934.493977950
224.491884.49188973950
234.49744.4974973967
244.500624.50062976972
254.515614.51561988975
264.497644.49764988975
274.512684.5074989980
284.506094.50062997985
294.512684.51561993983
304.512684.50764989985
314.494814.51668994995
324.514544.50609991989
334.524994.50268997991
344.502184.52268993992
354.530264.50481999997
364.517564.51454995994
374.506534.52499994996
384.502184.51218995991
394.485244.530261000997
404.517564.51756995992
414.494814.50653998990
424.4914.50218997991
434.4934.48524997996
444.491884.51756999990
454.49744.49481997997
464.500624.491999997
474.515614.493999997
484.497644.49188999997
494.512684.51268999995
504.50609999997
514.51268997998
524.512681000996
534.49481997994
544.514541000992
554.52499994998
564.5021810001000
574.53026998988
584.51756997996
594.50653999996
604.50218996999
614.48524999996
624.517561000994
634.49481997998
644.491999999
654.493998994
664.49188992994
674.4974998998
684.50062996999
694.515619991000
704.49764998996
714.51268996997
724.50609994994
734.51268999995
744.5126810001000
754.49481999994
764.51454999997
774.52499998998
784.50218985998
794.53026997982
804.517569971000
814.506539991000
824.50218998997
834.4852410001000
844.517569991000
854.494811000998
864.4919991000
874.4931000990
884.49188999996
894.4974999994
904.500621000998
914.515611000999
924.49764997999
934.512681000995
944.506099981000
954.51268998997
964.512681000999
974.49481999997
984.51454998996
994.52499997997
1004.502189961000
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
Sheet1
160
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.51268
Sheet2
192
4.4974
4.50062
4.51561
4.49764
4.51268
4.50609
4.51268
4.51268
4.49481
4.51454
4.52499
4.50218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.4974
4.50062
4.51561
4.49764
4.5074
4.50062
4.51561
4.50764
4.51668
4.50609
4.50268
4.52268
4.50481
4.51454
4.52499
4.51218
4.53026
4.51756
4.50653
4.50218
4.48524
4.51756
4.49481
4.491
4.493
4.49188
4.51268
Sheet3
160192224256160192
14.49744.4974004.49744.4974
24.500624.50062004.500624.50062
34.515614.51561004.515614.51561
44.497644.49764004.497644.49764
54.512684.51268514.512684.51268
64.506094.506092794.506094.50609
74.512684.5126871524.512684.51268
84.512684.51268171111double+add4.512684.51268
94.494814.494812712154.494814.49481
104.514544.514543613244.514544.51454
114.524994.524995074714.524994.52499
124.502184.502185945804.502184.50218
134.530264.530266566274.530264.53026
144.517564.517567536764.517564.51756
154.506534.506537887754.506534.50653
164.502184.502188508214.502184.50218
174.485244.485248788414.485244.48524
184.517564.517569219024.517564.51756
194.494814.494819319184.494814.49481
204.4914.4919579244.4914.491
214.4934.4939779504.4934.493
224.491884.491889739504.49744.49188
234.49744.4974973967double+double4.500624.4974
244.500624.500629769724.515614.50062
254.515614.515619889754.497644.51561
264.497644.497649889754.512684.49764
274.512684.50749899804.506094.5074
284.506094.500629979854.512684.50062
294.512684.515619939834.512684.51561
304.512684.507649899854.494814.50764
314.494814.516689949954.514544.51668
324.514544.506099919894.524994.50609
334.524994.502689979914.502184.50268
344.502184.522689939924.530264.52268
354.530264.504819999974.517564.50481
364.517564.514549959944.506534.51454
374.506534.524999949964.502184.52499
384.502184.512189959914.485244.51218
394.485244.5302610009974.517564.53026
404.517564.517569959924.494814.51756
414.494814.506539989904.4914.50653
424.4914.502189979914.4934.50218
434.4934.48524997996tau+add4.48524
444.491884.517569999904.51756
454.49744.494819979974.49481
464.500624.4919999974.491
474.515614.4939999974.493
484.497644.491889999974.49188
494.512684.512689999954.51268
504.50609999997
514.51268997998
524.512681000996
534.49481997994
544.514541000992
554.52499994998
564.5021810001000
574.53026998988
584.51756997996
594.50653999996
604.50218996999
614.48524999996
624.517561000994tau+tau
634.49481997998
644.491999999
654.493998994
664.49188992994
674.4974998998
684.50062996999
694.5156199910004.50218
704.497649989964.48524
714.512689969974.51756
724.506099949944.49481
734.512689999954.50218
744.51268100010004.48524
754.494819999944.51756
764.514549999974.49481
774.52499998998
784.50218985998
794.53026997982
804.517569971000
814.506539991000
824.502189989974.50218
834.48524100010004.48524
844.5175699910004.51756
854.4948110009984.49481
864.49199910004.493
874.49310009904.49188
884.491889999964.4974
894.49749999944.50062
904.5006210009984.51561
914.5156110009994.49764
924.497649979994.5074
934.5126810009954.50062
944.5060999810004.51561
954.512689989974.50764
964.5126810009994.51668
974.494819999974.50609
984.514549989964.50268
994.524999979974.52268
1004.5021899610004.50481
4.530264.51454
4.517564.52499
4.506534.51218
4.502184.53026
4.485244.51756
4.517564.50653
4.494814.50218
4.4914.48524
4.4934.51756
4.491884.49481
4.491
4.493
4.49188
4.51268
Sheet3
-
13
Details: Refer to Our Works
Tutorial on "Adversarial ML and Vehicular Networks: Strategies for Attack and Defense" at IEEE 89th Vehicular Technology Conference (VTC).
Selectd Research Papers DAC’19: "Building Robust Machine Learning Systems: Current
Progress, Research Challenges, and Opportunities". ICML’19 Workshop on Robust ML: “CapsAttacks: Robust and
Imperceptible Adversarial Attacks on Capsule Networks” DATE’19: “FAdeML: Understanding the Impact of Pre-
Processing Noise Filtering on Adversarial Machine Learning” IOLTS’19: “TrISec: Training Data-Unaware Imperceptible
Security Attacks on Deep Neural Networks”. IOLTS’19: “QuSecNets: Quantization-based Defense
Mechanism for Securing Deep Neural Network against Adversarial Attacks”.
-
14
Details: Refer to Our Works
Tutorial on "Adversarial ML and Vehicular Networks: Strategies for Attack and Defense" at IEEE 89th Vehicular Technology Conference (VTC).
Selectd Research Papers DAC’19: "Building Robust Machine Learning Systems: Current
Progress, Research Challenges, and Opportunities". ICML’19 Workshop on Robust ML: “CapsAttacks: Robust and
Imperceptible Adversarial Attacks on Capsule Networks” DATE’19: “FAdeML: Understanding the Impact of Pre-
Processing Noise Filtering on Adversarial Machine Learning” IOLTS’19: “TrISec: Training Data-Unaware Imperceptible
Security Attacks on Deep Neural Networks”. IOLTS’19: “QuSecNets: Quantization-based Defense
Mechanism for Securing Deep Neural Network against Adversarial Attacks”.
RTML Workshop at ITC 2019“Open for Submissions”
-
15
Data Poisoning Attacks during Training
T2’
T1
T2
T1 +T2’
T = T1 + T2
Trained Model
Training
InferenceStop
Attack 1: In this attack certain portion of the original dataset is intrudedwhich is then used for training.
60km/h
T = T1 + T2
T2
T1
T2’
T1 + T2 + T2’
Trained Model
Training
InferenceStop 60km/h
Attack 2: Instead of perturbing the original dataset, in this attack we extend thedataset with intruded samples which is then used for training.
Faiq, Shafique et al. @ FIT 2018, DSD’18, IOLTS 2019
-
16
Data Poisoning Attacks: Training16N0: No Noise N1: 1% Noise N2: .% Noise N1: 5% Noise
020406080
100
N0 N1 N3 N5
Top
Accu
racy
N0 N1 N3 N5
Stop
Yield
Dangerous Curve
Priority Road
Speed Limit 60
020406080
100
N0 N1 N3 N5
Top
Accu
racy
N0 N1 N3 N5
Turn Right
Dangerous Curve
Yield
Stop
Tun Left
Attack 1 Attack 2
Attack 1 Attack 2
Faiq, Shafique et al. @ FIT 2018, DSD’18, IOLTS 2019
-
17
Data Poisoning Attacks: Training17N0: No Noise N1: 1% Noise N2: .% Noise N1: 5% Noise
020406080
100
N0 N1 N3 N5
Top
Accu
racy
N0 N1 N3 N5
Stop
Yield
Dangerous Curve
Priority Road
Speed Limit 60
020406080
100
N0 N1 N3 N5
Top
Accu
racy
N0 N1 N3 N5
Turn Right
Dangerous Curve
Yield
Stop
Tun Left
Attack 1 Attack 2
Attack 1 Attack 2
Faiq, Shafique et al. @ FIT 2018, DSD’18, IOLTS 2019
To develop a successful (less impact on validation accuracy)Data Poisoning Attack, append the intruded samples in original
dataset instead of changing the original dataset
Limitations:1. Require the access to the complete training dataset2. The attack Noise is perceptible.
-
18
Adversarial Attacks on ML-Systems
Specially crafted imperceptible noise to misclassify the ML System Imperceptibility Robustness
Output Label = Speed Limit
(60km/h)
Buffer
Preprocessing Noise filters
DNN
Input Label = Stop
Perturbed Image
Environmental Noise and Variations
Challenges• Ensure the imperceptibility• Handling the impact of environmental Noise
Subjective Analysis
Faiq, Shafique et al. @ IOLTS 2019
-
19
Adversarial Attacks
DNN
Training Dataset
DNN
Trained Model
Network Parameters
W11W21
Wn1
W12W22
Wn2
W1nW2n
Wnn
b1b2
bn
Gradient-basedAdversarial Attacks
(a) White-Box Attacks
Adversarial Example
DN
N Architecture
Training
(b) Black-Box Attacks
DNN
Training Dataset
DNN
Trained Model
Probability vectorScore-based
Decision-based Output Label
Model StealingDNN
Network Parameters
W11
W21
Wn1
W12
W22
Wn2
W1n
W2n
Wnn
b1
b2
bn
Gradient-basedAdversarial Attacks
Adversarial Example
Adversarial Example
Transfer Attacks
DNN Architecture
Training
Faiq, Shafique et al. @ IOLTS 2019
-
20
TrISec: Training Data Unaware Adversarial Attack
Deep Neural Network
Perturbed Image(x+δ)
Cost Functi
on
𝐶𝐶 = �𝑗𝑗=0
𝑛𝑛𝐿𝐿−1
(𝑎𝑎𝑗𝑗𝐿𝐿 − 𝑦𝑦𝑗𝑗)2
Current Distribution
Targeted Distribution
Compute Correlation
(CR)
PASS
𝜕𝜕𝐶𝐶
𝜕𝜕𝑎𝑎𝑗𝑗(𝐿𝐿) = 2 �
𝑗𝑗=0
𝑛𝑛𝐿𝐿−1
(𝑎𝑎𝑗𝑗𝐿𝐿 − 𝑦𝑦𝑗𝑗)
𝜕𝜕𝐶𝐶
𝜕𝜕𝑎𝑎𝑗𝑗(𝐿𝐿−1) = �
𝑗𝑗=0
𝑛𝑛𝐿𝐿−1
𝑤𝑤𝑗𝑗𝑗𝑗𝐿𝐿 𝜎𝜎𝜎 𝑧𝑧𝑗𝑗
𝐿𝐿−1 𝜕𝜕𝐶𝐶
𝜕𝜕𝑎𝑎𝑗𝑗(𝐿𝐿)
Compute the effect of cost (C) on activation function (a)
FAIL
Compute structural similarity index (SS)
Original Image
(x)
Attack Image
PASS FAIL
PASS
0
50
100
A B C D E
0
50
100
A B C D E
Targeted Class or Targeted Classification
Probability
FAIL
Update Noise /
Perturbation (𝛿𝛿)
𝑎𝑎(𝐿𝐿)
𝑦𝑦
0.95 ≤ 𝐶𝐶𝐶𝐶 ≤ 1 𝛿𝛿 = 𝛿𝛿 × (1 − 𝐶𝐶𝐶𝐶)
𝛿𝛿 = 𝛿𝛿 × (1 − 𝑆𝑆𝑆𝑆)
C ≤ ε
0.99 ≤ 𝑆𝑆𝑆𝑆 ≤ 1
F. Khalid, M. A. Hanif, S. Rehman, M. Shafique, “TrISec: Training Data-Unaware Imperceptible Security Attacks on Machine Learning Modules of Autonomous Vehicles”, IOLTS2019. [Link: https://arxiv.org/abs/1811.01031]
+
Perturbation (𝛿𝛿)
G2: Imperceptibility
G1: Misclassification
https://arxiv.org/abs/1811.01031
-
21
Imperceptible Data Poisoning Attacks during Inference
Key Insights:• Instead of training data samples, the probability vector at the
output can be used to optimize the attack noise.• To ensure the imperceptibility, additional similarity and correlation
checks are required
I0: No Noise I1: 1st Iteration I25: 25Kth Iteration I50: 50Kth Iteration
0.80.850.90.951
0255075
100
I0 I1 I25 I50
SSI,
CR
Top
5 Ac
cura
cy
StopYieldDangerous CurvePriority RoadSpeed Limit 60CRSSI
Top 5 Accuracy, Correlation
Coefficient (CR) and Structural Similarity
Index (SSI)
F. Khalid, M. A. Hanif, S. Rehman, M. Shafique, “TrISec: Training Data-Unaware Imperceptible Security Attacks on Machine Learning Modules of Autonomous Vehicles”, IOLTS 2019. [Link: https://arxiv.org/abs/1811.01031 ]
https://arxiv.org/abs/1811.01031
-
22
Practical Implications of Adversarial Perturbations22
Input Label = Stop Output Label =
Speed Limit (60km/h)
Preprocessing Noise filters
Perturbed Image
Attacker can physically insert the perturbations in sensor’s data
DNN 020406080
100
A B C D E
Buffe
r
Subjective Analysis
Input Label = Stop Output Label =
Speed Limit (60km/h)
Preprocessing Noise filters
Perturbed Image
Attacker has access to the output of the pre-processing module
DNN 020406080
100
A B C D E
Buffe
r
Subjective Analysis
Faiq, Shafique et al. @ DATE 2019
Input Label = Stop Output Label =
Speed Limit (60km/h)
Preprocessing Noise filters
Perturbed Image
Attacker has access to the input of the pre-processing module
DNN 020406080
100
A B C D E
Buffe
r
Subjective Analysis
-
23
Potential Defenses: FAdeML
Cost Function 𝑓𝑓(𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐)
Buffer
Preprocessing Noise Filters
0
30
60
90
A B C D E
020406080
100
A B C D E
Adversarial Attack LibraryI-FGSM, JSMA, FGSM, TrISec
Optimization AlgorithmSpecific to selected
Adversarial Attack
Cost Function 𝑓𝑓(𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐)
Input samples
Adversarial Examples
Reference Sample(s)
020406080
100
A B C D E
020406080
100
A B C D E
75.68%
Stop Stop Sign
Stop Stop Sign
Stop 60km/h
Stop 60km/h
FGSM I-FGSM JSMA
85.61% 89.24% 92.58%
TrISec
78.45% 72.14% 62.15% 68.68%
68.45% 72.36% 74.28% 75.12%
99.47% 99.47% 99.47% 99.47%DNN
DNN
DNN
DNN
Analysis 1: Attacker has direct access to the input of DNN, i.e., access to the output of
the preprocessing module.
Analysis 2: Attacker does not direct access to the input of DNN and does not have knowledge of preprocessing module.
Analysis 3: Attacker does not direct access to the input of DNN and have the complete
knowledge of the preprocessing module.Faiq, Shafique et al. @ DATE 2019
-
24
Defenses Against Adversarial ExamplesGoal Increase the perceptibility of the attack noise
Decrease the robustness of the attack, i.e., nullify the effects of imperceptible attack Noise
Deer
Airplane Automobile
Deer
Before Defense After Defense
Before Defense After Defense
Automobile
Automobile
Original
Original
Faiq, Shafique et al. @ IOLTS 2019
-
25
IOLTS’19: Effects of Quantization on Adversarial ML25
A
A
A
Correct Classification of Clean ImagesIncorrect Classification of Perturbed Images
Correct Classification of Perturbed Images
Clean Images
Perturbed Image
Quantized Clean ImagesQuantized Perturbed ImageNormal Process
Adversarial Attacks
Claverhans(FGSM, CW,
JSMA) CNNQuantization Layer (QL)
CNN
CIFAR-10
Perturbed Image Perturbed Quantized
Images
P(Ship) = 0.9037
Clean Image
P(Airplane) = 0.7969
P(Airplane) = 0.6987
P(Airplane) = 0.8372
Clean Quantized
Images
P(3) = 0.862
P(8) = 0.457
P(1) = 1.0
P(3) = 0.911
P(8) = 0.524
P(3) = 0.897
P(Dog) = 0.75
P(Automobile) = 0.91
P(Deer) = 0.97
P(Dog) = 0.69
P(Dog) = 0.34
P(3) = 0.55
WithQuantization Layer
Without Quantization Layer
WithQuantization Layer
Without Quantization Layer
No Attack
FGSM Attack
C&W Attack
-
26
IOLTS’19: Effects of Quantization on Adversarial ML26
A
A
A
Correct Classification of Clean ImagesIncorrect Classification of Perturbed Images
Correct Classification of Perturbed Images
Clean Images
Perturbed Image
Quantized Clean ImagesQuantized Perturbed ImageNormal Process
Adversarial Attacks
Claverhans(FGSM, CW,
JSMA) CNNQuantization Layer (QL)
CNN
CIFAR-10
Perturbed Image Perturbed Quantized
Images
P(Ship) = 0.9037
Clean Image
P(Airplane) = 0.7969
P(Airplane) = 0.6987
P(Airplane) = 0.8372
Clean Quantized
Images
P(3) = 0.862
P(8) = 0.457
P(1) = 1.0
P(3) = 0.911
P(8) = 0.524
P(3) = 0.897
P(Dog) = 0.75
P(Automobile) = 0.91
P(Deer) = 0.97
P(Dog) = 0.69
P(Dog) = 0.34
P(3) = 0.55
WithQuantization Layer
Without Quantization Layer
WithQuantization Layer
Without Quantization Layer
No Attack
FGSM Attack
C&W Attack
Challenging Question
Can quantization as an input transformation be used to increase the perceptibility of the adversarial noise
while preserving the classification accuracy?
-
27
IOLTS’19: QuSecNets – Quantization-based Defense 27
Constant
Trainable
Quantization
QL
TrainingOnly for CQ Only for TQ
For Both CQ and TQ
QuSecNets
Training Dataset
DNN Model
QL: Quantization Layer
n = Number of Quantization Levels𝑐𝑐𝑗𝑗 is trainable for Trainable quantization
+
𝑆𝑆(𝑥𝑥, 𝑐𝑐1)
𝑆𝑆(𝑥𝑥, 𝑐𝑐2)
𝑆𝑆(𝑥𝑥, 𝑐𝑐𝑛𝑛−1)
𝑦𝑦1
𝑦𝑦2
𝑦𝑦𝑛𝑛−1
𝑥𝑥
1𝑛𝑛 − 1
𝑦𝑦
Error Injection
Error Resilience Analysis
Selecting “n”
Maximum Imperceptible Perturbations
Maximum Tolerable Noise
Faiq, Shafique et al. @ IOLTS 2019
-
28
Experimental Results28
92
94
96
98
100
0 0.1 0.2 0.3
Q_2
Q_3
Q_4
Q_5
Q_6Maximum allowed perturbations in a single pixel (epsilon)
Clas
sific
atio
n A
ccur
acy
unde
r Att
ack
Scen
ario Trainable
QuantizationConstant Quantization
Maximum allowed perturbations in a single pixel (epsilon)
Clas
sific
atio
n Ac
cura
cy
unde
r Att
ack
Scen
ario
0
50
100
0 0.1 0.2 0.3
No Defense
Bird Deer
FrogShip
Airplane
Airplane
Automobile
Trainable Quantization
Constant Quantization
No Defense
Trainable Quantization performs better than Constant Quantization
Increase the perceptibility of the adversarial noise
Increase the classification accuracy under attack
FSGM Attack
JSMA Attack
Faiq, Shafique et al. @ IOLTS 2019
Maximum allowed perturbations in a single pixel (epsilon)
Classification Accuracy under Attack Scenario
1
Q_200.10.20.398.898.6798.598.5Q_300.10.20.399.198.8398.2597.67Q_400.10.20.399.398.6397.4996Q_500.10.20.399.398.596.6894.61Q_600.10.20.399.398.2594.7192.69
Trainable Quantization
Constant Quantization
Maximum allowed perturbations in a single pixel (epsilon)
Classification Accuracy under Attack Scenario
No Defense
00.10.20.310010.521.25000.10.20.399.297.1194.792.1800.10.20.398.9798.4697.7797.1
-
29
The imperceptibility of the adversarial noise increases with the increase in number of the allowed queries.
Input Label = Stop
Preprocessing
DNNP(1)P(2)P(3)
P(n)
Probability Vector
Black-boxP(1)
Top-1 Probability
P(1)
Top-1 Label
Decision–based Attacks
Clean ImagesCIFAR-10 Dataset
Adversarial examples when number of queries (Q) is
restricted to 1000
Adversarial examples when number of queries (Q) is
restricted to 10000
Case I Case II Case III Case IV Case V
0
50
100
150
200
250
300
Case I Case II Case III Case IV Case V
Pertu
rbat
ion
Nor
m
Q = 1000 Q = 10000
94.9
275.
8
293.
1
269.
2
1.51
2.44
1.54
3.47
0.7913
Decision-based Adversarial Attacks
-
30
The imperceptibility of the adversarial noise increases with the increase in number of the allowed queries.
Input Label = Stop
Preprocessing
DNNP(1)P(2)P(3)
P(n)
Probability Vector
Black-boxP(1)
Top-1 Probability
P(1)
Top-1 Label
Decision–based Attacks
Clean ImagesCIFAR-10 Dataset
Adversarial examples when number of queries (Q) is
restricted to 1000
Adversarial examples when number of queries (Q) is
restricted to 10000
Case I Case II Case III Case IV Case V
0
50
100
150
200
250
300
Case I Case II Case III Case IV Case V
Pertu
rbat
ion
Nor
m
Q = 1000 Q = 10000
94.9
275.
8
293.
1
269.
2
1.51
2.44
1.54
3.47
0.7913
Challenging Question
Can new decision-based attacks be devised that can operate successfully in real-time resource-
constrained applications while ensuring imperceptibility and robustness of attacks?
Decision-based Adversarial Attacks
-
31
The imperceptibility of the adversarial noise increases with the increase in number of the allowed queries.
Input Label = Stop
Preprocessing
DNNP(1)P(2)P(3)
P(n)
Probability Vector
Black-boxP(1)
Top-1 Probability
P(1)
Top-1 Label
Decision–based Attacks
Clean ImagesCIFAR-10 Dataset
Adversarial examples when number of queries (Q) is
restricted to 1000
Adversarial examples when number of queries (Q) is
restricted to 10000
Case I Case II Case III Case IV Case V
0
50
100
150
200
250
300
Case I Case II Case III Case IV Case V
Pertu
rbat
ion
Nor
m
Q = 1000 Q = 10000
94.9
275.
8
293.
1
269.
2
1.51
2.44
1.54
3.47
0.7913
Challenging Question
Can new decision-based attacks be devised that can operate successfully in real-time resource-
constrained applications while ensuring imperceptibility and robustness of attacks?
Decision-based Adversarial Attacks
F. Khalid, M. Shafique et al., “RED-Attack: Resource Efficient Decision based Attack
for Machine Learning”, [Link: https://arxiv.org/abs/1901.10258]
-
32
My Research Team and CollaboratorsPost-Docs and PhDs
Collaborators
32
MS/BS Students
Previous Students
-
Computer Architecture and RobustEnergy-Efficient Technologies Group
CAREech
Thank you!Questions?
Security for Machine Learning Systems �Attacks, Defenses, & ReconfigurabilityWho Ruled the World!Smart Cyber Physical Systems & Internet-of-ThingsSmart CPS & IoT => The Big Data Processing Challenge!Smart CPS & IoT => The Big Data Processing Challenge!ML Applications => requiring High Efficiency GainsMachine Learning and the Computing Efficiency Gap!Machine Learning and the Computing Efficiency Gap!Brain-Inspired Computing: [email protected] for Machine Learning: News FeedRisks that Arise from Adversarial ExamplesReliability and Security for Machine Learning SystemsDetails: Refer to Our WorksDetails: Refer to Our WorksData Poisoning Attacks during TrainingData Poisoning Attacks: TrainingData Poisoning Attacks: TrainingAdversarial Attacks on ML-SystemsAdversarial AttacksTrISec: Training Data Unaware Adversarial AttackImperceptible Data Poisoning Attacks during InferencePractical Implications of Adversarial PerturbationsPotential Defenses: FAdeMLDefenses Against Adversarial ExamplesIOLTS’19: Effects of Quantization on Adversarial MLIOLTS’19: Effects of Quantization on Adversarial MLIOLTS’19: QuSecNets – Quantization-based Defense Experimental ResultsDecision-based Adversarial AttacksDecision-based Adversarial AttacksDecision-based Adversarial AttacksMy Research Team and CollaboratorsSlide Number 33