Home

Awesome

Towards an Awareness of Time Series Anomaly Detection Models’ Adversarial Vulnerability

Overview

Pipeline of a typical time series training phase and adversarial attack phase

<img src="https://i.ibb.co/WfLfgMS/Adv-ICLR-Pipeline.png" alt="Coffee" border="0"> <hr/>

Example of ground truth vs perturbed time series using FGSM and PGD attacks on CLMPPCA

<img src="https://i.ibb.co/NW9CtM1/Realvs-Adv.png" alt="Coffee" border="0" width="500" ><img src="https://i.ibb.co/pJ7mfmX/Realvs-Adv-error.png" alt="Coffee" border="0" width="500" >

<hr/>

FGSM and PGD attacks on MSCRED (left) and MTAD-GAT (right)

<img src="https://i.ibb.co/Sm4SpFy/MSCRED.png" alt="Coffee" border="0" width="500" ><img src="https://i.ibb.co/j5LMBYB/MTAD.png" alt="Coffee" border="0" width="500" >

<hr/> <!--- <img src="https://i.ibb.co/87Fxy61/MSCRED-3.png" alt="Coffee" border="0" width="500" > -->

Additional Experiments on UCR Dataset

In addition to all the experiment on state-of-the-art anomaly and intrusion detection system. We also cover general time series classification task where we attack a multilayer perception (MLP), a fully convolutional network and ResNet trained on different dataset from the UCR repository. We conduct an analysis of 71 datasets from the University of California, Riverside (UCR) repository. In future work, we will expand on this experiment by including additional neural networks (MobileNet, EfficientNet, DenseNet, and Inception Time) and datasets (the remainder of the UCR dataset, datasets from the UEA repository).

We find that the Carlini-Wagner L2 attack provides the best adversarial examples, resulting in the most significant performance degradation. In Figure below, we show some original samples and the corresponding perturbed samples generated by FGSM, PGD, BIM, Carlini-Wagner L2, and MIM attacks on UCR datasets. Also, in Tables below, we present the classification results for MLP, FCN, and ResNet.

<img src="https://i.ibb.co/bRWfBNJ/Adiac-attacks-2.png" alt="Adiac" border="0" width="500"><img src="https://i.ibb.co/QKnpsZh/Car-attacks-1.png" alt="Car" border="0" width="500" > <img src="https://i.ibb.co/f1cKd6z/FISH-attacks-0.png" alt="FISH" border="0" width="500"><img src="https://i.ibb.co/KDzGtwW/Meat-attacks-2.png" alt="Meat" border="0" width="500"> <img src="https://i.ibb.co/SyR00sK/Coffee-attacks-1.png" alt="Coffee" border="0" width="500"><img src="https://i.ibb.co/RCHbRQJ/Diatom-Size-Reduction-attacks-0.png" alt="Diatom-Size-Reduction" border="0" width="500">


Adversarial attacks on an MLP trained on different UCR datasets

DatasetsFGSMPGDBIMMadry et al. 2016Carlini <br> Wagner <br> L2MIMNo <br> Attack
50words44±0.842±1.342±1.342±1.335±143±163±1.1
Adiac14±1.815±1.615±1.615±1.616±1.316±1.853±2.7
ArrowHead29±3.927±3.327±3.327±3.324±4.327±3.174±2.6
Beef32±5.126±3.926±3.926±3.929±3.927±3.478±3.9
BeetleFly74±7.774±7.774±7.774±7.770±574±7.775±13.3
BirdChicken62±5.862±5.862±5.862±5.857±10.562±5.869±2.9
Car49±135±2.935±2.935±2.952±145±183±1
CBF76±2.676±2.476±2.476±2.463±3.576±2.694±2.6
Chlorine <br> Concentration24±0.324±0.524±0.524±0.524±0.724±0.465±0.4
Coffee9±2.19±2.19±2.19±2.19±4.29±2.1100±0
Computers46±1.145±1.145±1.145±1.145±1.145±1.158±0.9
Cricket_X26±0.725±0.725±0.725±0.721±126±0.945±1
Cricket_Y30±0.829±1.729±1.729±1.724±0.629±1.748±1.6
Cricket_Z32±0.730±1.130±1.130±1.125±131±0.244±1.2
DiatomSize <br> Reduction40±1.237±1.437±1.437±1.431±438±1.595±2.4
DistalPhalanx <br> OutlineAgeGroup16±0.916±116±116±116±116±0.983±0.8
DistalPhalanx <br> OutlineCorrect29±1.428±1.828±1.828±1.825±0.929±1.777±0.9
Distal <br> PhalanxTW13±0.812±1.112±1.112±1.112±0.912±1.278±0.7
Earthquakes69±1.569±1.569±1.569±1.552±4.269±1.573±1.1
ECG20060±1.860±2.160±2.160±2.129±5.860±2.184±0.6
ECG500065±0.264±0.364±0.364±0.361±0.364±0.493±0.2
ECGFiveDays48±2.346±2.146±2.146±2.135±4.847±2.195±3.3
ElectricDevices22±0.421±0.521±0.521±0.521±0.621±0.655±0.8
FaceAll57±0.356±0.456±0.456±0.439±0.956±0.274±0.6
FaceFour79±2.477±277±277±276±1.879±1.488±0.7
FacesUCR67±1.763±1.663±1.663±1.655±1.665±1.883±1.2
FISH16±2.18±1.28±1.28±1.214±1.912±1.285±0.4
Gun_Point48±6.147±6.247±6.247±6.234±5.447±6.292±1.4
Ham34±2.434±2.634±2.634±2.648±3.534±2.670±2
Haptics21±0.921±0.821±0.821±0.821±1.220±0.741±0.7
Herring50±1.950±1.950±1.950±1.950±1.950±1.951±1.9
InlineSkate21±1.119±0.819±0.819±0.820±1.420±1.434±0.7
InsectWingbeat <br> Sound37±0.730±0.330±0.330±0.342±0.334±0.462±0.7
ItalyPower <br> Demand82±0.882±0.982±0.982±0.911±1.482±0.996±0.2
LargeKitchen <br> Appliances33±2.232±1.332±1.332±1.334±0.633±2.151±0.5
Lighting270±2.670±2.670±2.670±2.658±3.870±2.665±3.5
Lighting753±4.253±3.753±3.753±3.735±3.753±3.764±2.4
Meat26±126±126±126±125±1.726±174±1
MedicalImages39±1.936±2.236±2.236±2.226±0.537±2.267±0.5
MiddlePhalanx <br> OutlineAgeGroup32±10.726±4.826±4.826±4.820±0.827±5.773±1.5
MiddlePhalanx <br> OutlineCorrect46±1.546±1.646±1.646±1.645±1.546±1.656±1.5
Middle <br> PhalanxTW18±2.918±2.818±2.818±2.818±1.718±2.956±2.4
MoteStrain79±0.779±0.779±0.779±0.753±2.379±0.784±1.1
OliveOil28±228±228±228±228±228±259±2
OSULeaf29±0.729±1.129±1.129±1.129±0.930±0.745±0.3
Phalanges <br> OutlinesCorrect33±3.233±2.633±2.633±2.633±2.333±2.768±2.4
Plane89±287±1.187±1.187±1.185±4.388±1.198±1.1
ProximalPhalanx <br> OutlineAgeGroup18±218±2.318±2.318±2.318±1.818±2.381±1.9
ProximalPhalanx <br> OutlineCorrect36±1.434±1.134±1.134±1.133±1.634±0.968±1.6
Proximal <br> PhalanxTW41±3.942±442±442±442±442±3.953±4.1
Refrigeration <br> Devices36±1.836±1.636±1.636±1.636±1.336±1.943±1.2
ScreenType39±1.438±1.838±1.838±1.838±139±1.636±0.3
ShapeletSim50±1.750±1.450±1.450±1.449±1.750±1.448±0.9
ShapesAll49±1.642±1.142±1.142±1.143±1.346±1.870±0.2
SmallKitchen <br> Appliances33±1.434±134±134±136±1.634±1.149±2.2
SonyAIBO <br> RobotSurface68±2.668±2.668±2.668±2.662±7.368±2.668±4.6
SonyAIBO <br> RobotSurfaceII81±0.881±0.881±0.881±0.871±0.681±0.883±0.8
Strawberry7±0.36±0.36±0.36±0.39±0.77±0.296±0.3
SwedishLeaf32±1.226±2.126±2.126±2.125±0.829±1.482±0.3
Symbols76±1.574±1.274±1.274±1.276±1.475±189±0.2
synthetic_control80±1.680±1.780±1.780±1.737±3.680±1.695±1
ToeSegmentation151±1.551±1.551±1.551±1.550±1.251±1.557±0.7
ToeSegmentation263±1.863±1.863±1.863±1.855±5.563±1.867±3
Trace29±2.729±2.429±2.429±2.429±2.429±2.989±1.8
TwoLeadECG45±2.244±2.344±2.344±2.337±1.845±2.277±0.7
Two_Patterns32±1.831±1.631±1.631±1.612±0.231±1.796±0.4
wafer39±1.539±1.539±1.539±1.521±1.539±1.596±0.9
Wine45±045±045±045±045±045±056±0
WordsSynonyms40±1.238±0.538±0.538±0.532±139±1.153±0.4
Worms28±0.427±0.927±0.927±0.924±1.528±0.636±1.2
WormsTwoClass49±1.249±149±149±147±1.449±160±1

Adversarial attacks on a FCN trained on different UCR datasets

DatasetsFGSMPGDBIMMadry et al. 2016Carlini Wagner L2MIMNo Attack
50words3±0.56±1.46±1.46±1.418±3.64±1.329±16
Adiac5±1.87±3.87±3.87±3.811±2.17±3.524±17.7
ArrowHead40±014±6.214±6.214±6.214±6.515±680±6.6
Beef26±10.223±9.723±9.723±9.723±12.722±7.752±9.7
BeetleFly50±020±520±520±520±520±580±5
BirdChicken50±015±1015±1015±107±2.922±2.994±2.9
Car22±040±27.540±27.540±27.540±26.240±25.147±23.4
CBF83±1.279±1.679±1.679±1.61±0.181±1.3100±0.2
Chlorine <br> Concentration39±19.539±19.839±19.839±19.838±19.139±19.854±18.5
Coffee0±00±00±00±00±00±0100±0
Computers44±1019±5.719±5.719±5.716±6.128±1185±6.1
Cricket_X16±5.711±1.811±1.811±1.813±2.311±372±3.7
Cricket_Y19±1.916±3.116±3.116±3.116±2.916±3.369±7.5
Cricket_Z13±1.111±3.211±3.211±3.214±3.511±2.172±5.1
DiatomSize <br> Reduction16±4.96±0.96±0.96±0.97±0.57±0.793±0.7
DistalPhalanx <br> OutlineAgeGroup19±4.719±4.419±4.419±4.419±4.419±4.480±4.3
DistalPhalanx <br> OutlineCorrect38±9.632±6.132±6.132±6.132±6.233±6.669±6.1
Distal <br> PhalanxTW15±1.117±1.217±1.217±1.217±1.117±1.173±2.1
Earthquakes36±4.134±3.234±3.234±3.225±2.535±3.376±2.5
ECG20049±6.516±3.116±3.116±3.111±1.824±589±1.8
ECG500069±6.933±24.733±24.733±24.74±0.451±12.594±0.4
ECGFiveDays38±9.52±0.22±0.22±0.22±0.32±0.399±0.3
ElectricDevices43±1.332±2.732±2.732±2.714±3.335±2.970±3.7
FaceAll66±0.741±0.441±0.441±0.48±2.757±0.490±2.8
FaceFour6±2.33±1.83±1.83±1.85±1.83±1.294±0.7
FacesUCR68±2.440±7.940±7.940±7.94±0.756±4.493±0.8
FISH13±0.419±11.519±11.519±11.522±11.918±1160±2.9
Gun_Point51±2.72±0.72±0.72±0.71±0.44±2.4100±0.4
Ham37±3.437±3.537±3.537±3.537±3.537±3.564±3.5
Haptics23±3.118±4.818±4.818±4.819±518±4.829±3.4
Herring60±046±8.246±8.246±8.249±11.954±5.560±0
InlineSkate16±0.513±5.213±5.213±5.216±6.713±4.522±7.6
InsectWingbeat <br> Sound13±1.811±1.311±1.311±1.312±1.511±1.423±4.4
ItalyPower <br> Demand84±181±1.781±1.781±1.75±0.583±1.596±0.3
LargeKitchen <br> Appliances50±4.932±23.732±23.732±23.721±17.545±13.974±16
Lighting240±1.729±129±129±129±130±1.772±1
Lighting732±7.619±2.919±2.919±2.917±3.523±4.274±1.6
Meat34±045±13.745±13.745±13.752±24.947±11.734±0
MedicalImages23±6.814±214±214±214±3.116±1.277±2.8
MiddlePhalanx <br> OutlineAgeGroup18±6.618±5.918±5.918±5.917±5.718±6.170±6.7
MiddlePhalanx <br> OutlineCorrect44±22.543±21.643±21.643±21.645±24.243±21.658±21.4
Middle <br> PhalanxTW20±1023±1123±1123±1121±923±10.748±12.8
MoteStrain80±178±1.278±1.278±1.210±0.579±1.591±0.5
OliveOil18±19.316±21.216±21.216±21.218±19.318±19.356±15.1
OSULeaf14±012±412±412±412±4.411±4.175±16.7
Phalanges <br> OutlinesCorrect36±2.536±2.536±2.536±2.536±2.636±2.565±2.6
Plane40±5.811±3.911±3.911±3.90±025±6.5100±0
ProximalPhalanx <br> OutlineAgeGroup32±23.722±8.822±8.822±8.825±10.722±8.864±18.9
ProximalPhalanx <br> OutlineCorrect32±26.831±26.431±26.431±26.431±26.231±26.870±26.2
Proximal <br> PhalanxTW18±8.214±3.114±3.114±3.115±4.714±2.975±2.9
Refrigeration <br> Devices40±3.536±0.936±0.936±0.935±1.736±146±1.7
ScreenType33±3.328±3.628±3.628±3.627±3.629±4.362±5.2
ShapeletSim8±3.78±3.18±3.18±3.18±2.88±3.193±2.8
ShapesAll4±1.43±2.93±2.93±2.97±0.63±1.919±18
SmallKitchen <br> Appliances53±16.737±18.137±18.137±18.139±22.641±11.143±12.3
SonyAIBO <br> RobotSurface84±2.282±2.782±2.782±2.75±0.383±2.797±0.6
SonyAIBO <br> RobotSurfaceII86±1.584±2.184±2.184±2.13±0.585±1.798±0.5
Strawberry44±20.831±8.831±8.831±8.831±8.931±9.170±8.8
SwedishLeaf28±1.710±2.610±2.610±2.66±3.613±3.393±3.6
Symbols36±3.26±1.66±1.66±1.65±0.615±1.994±1.3
synthetic_control95±195±1.395±1.395±1.33±0.995±1.298±0.7
ToeSegmentation141±6.211±0.811±0.811±0.83±0.718±398±0.7
ToeSegmentation243±1.426±2.326±2.326±2.314±2.836±0.587±2.8
Trace52±18.618±8.918±8.918±8.91±0.643±2.9100±0.6
TwoLeadECG7±3.12±0.42±0.42±0.41±0.13±0.7100±0.1
Two_Patterns34±7.315±0.715±0.715±0.715±0.719±2.386±0.7
wafer8±3.23±0.93±0.93±0.91±0.23±1.3100±0.2
Wine50±050±050±050±050±050±050±0
WordsSynonyms5±2.29±3.39±3.39±3.312±1.56±1.930±10.2
Worms17±1.721±3.621±3.621±3.621±5.321±3.448±7.3
WormsTwoClass48±539±2.339±2.339±2.339±2.540±4.262±2.3

Adversarial attacks on ResNet trained on different UCR datasets

DatasetsFGSMPGDBIMMadry et al. 2016Carlini Wagner L2MIMNo Attack
50words8±2.310±110±110±113±1.59±1.567±0.7
Adiac5±0.210±1.210±1.210±1.210±0.210±0.482±0.7
ArrowHead34±11.513±0.913±0.913±0.913±1.515±179±2.3
Beef24±8.919±5.119±5.119±5.118±3.922±3.974±3.4
BeetleFly29±5.817±5.817±5.817±5.817±5.817±5.884±5.8
BirdChicken54±5.814±2.914±2.914±2.914±2.920±587±2.9
Car20±19±4.59±4.59±4.58±3.910±4.989±3.5
CBF89±1.487±1.887±1.887±1.81±0.288±1.6100±0.2
Chlorine <br> Concentration14±0.414±0.814±0.814±0.813±0.414±0.782±1.1
Coffee0±00±00±00±00±00±0100±0
Computers58±5.424±1.324±1.324±1.320±3.245±5.182±2.6
Cricket_X33±317±2.517±2.517±2.514±2.127±1.976±2.4
Cricket_Y23±0.613±0.713±0.713±0.713±0.616±1.780±1.1
Cricket_Z28±2.914±214±214±213±0.822±2.478±1.4
DiatomSize <br> Reduction10±4.14±1.54±1.54±1.55±24±1.497±1.9
DistalPhalanx <br> OutlineAgeGroup18±2.417±1.817±1.817±1.817±217±1.881±1.8
DistalPhalanx <br> OutlineCorrect29±3.623±123±123±121±1.225±1.780±1
Distal <br> PhalanxTW15±0.315±0.815±0.815±0.814±0.615±0.976±0.7
Earthquakes48±2.945±2.745±2.745±2.724±146±3.180±1.2
ECG20069±4.450±11.650±11.650±11.613±2.163±4.188±2.4
ECG500073±0.861±1.361±1.361±1.35±0.366±1.394±0.3
ECGFiveDays33±16.24±1.64±1.64±1.63±0.66±3.898±0.7
ElectricDevices41±2.131±1.731±1.731±1.715±2.436±2.270±4.5
FaceAll76±0.469±169±169±111±0.574±0.783±1.6
FaceFour30±5.29±2.49±2.49±2.44±2.922±3.595±0.7
FacesUCR74±1.464±2.364±2.364±2.33±0.870±1.695±0.4
FISH13±03±0.93±0.93±0.93±1.23±0.998±1
Gun_Point23±5.66±26±26±21±0.410±0.7100±0
Ham30±2.929±229±229±230±2.429±272±2
Haptics20±0.222±3.122±3.122±3.121±3.721±3.849±4
Herring49±1141±141±141±141±141±160±1
InlineSkate15±1.319±219±219±219±2.919±1.932±3.1
InsectWingbeat <br> Sound22±0.523±0.623±0.623±0.623±0.424±0.346±1.1
ItalyPower <br> Demand87±1.386±0.886±0.886±0.87±0.986±1.397±0.2
LargeKitchen <br> Appliances59±2.832±2.732±2.732±2.78±1.447±1.290±0.8
Lighting246±042±2.642±2.642±2.627±1.743±1.774±1.7
Lighting736±3.720±4.220±4.220±4.219±2.124±7.774±4.2
Meat17±15.58±5.48±5.48±5.48±5.48±5.493±5.4
MedicalImages47±528±3.828±3.828±3.815±2.536±2.478±0.7
MiddlePhalanx <br> OutlineAgeGroup16±1.516±0.716±0.716±0.715±0.216±0.875±1
MiddlePhalanx <br> OutlineCorrect27±9.127±927±927±927±9.127±974±9.2
Middle <br> PhalanxTW15±2.817±0.417±0.417±0.417±0.617±0.762±0.8
MoteStrain76±0.973±1.173±1.173±1.110±0.875±1.191±0.8
OliveOil14±017±5.817±5.817±5.818±3.917±5.879±2
OSULeaf14±16±2.26±2.26±2.25±1.96±2.294±2.8
Phalanges <br> OutlinesCorrect27±317±0.917±0.917±0.918±0.717±0.984±0.9
Plane73±6.241±6.441±6.441±6.40±063±5.3100±0
ProximalPhalanx <br> OutlineAgeGroup16±4.815±0.815±0.815±0.816±1.515±0.886±0.6
ProximalPhalanx <br> OutlineCorrect16±2.611±1.611±1.611±1.611±1.711±1.690±1.6
Proximal <br> PhalanxTW8±1.213±0.513±0.513±0.514±0.314±0.482±0.5
Refrigeration <br> Devices35±2.534±3.134±3.134±3.131±2.334±3.154±0.6
ScreenType35±729±2.629±2.629±2.628±3.532±4.561±3.8
ShapeletSim13±7.912±8.612±8.612±8.610±10.213±8.191±9.9
ShapesAll7±0.73±0.33±0.33±0.35±0.34±0.788±0.5
SmallKitchen <br> Appliances44±4.528±5.628±5.628±5.629±7.734±5.256±16
SonyAIBO <br> RobotSurface80±2.379±2.979±2.979±2.914±3.279±2.592±0.9
SonyAIBO <br> RobotSurfaceII81±1.179±1.679±1.679±1.64±0.880±198±0.8
Strawberry24±1622±17.722±17.722±17.722±17.622±17.780±17.6
SwedishLeaf34±0.816±0.516±0.516±0.54±0.522±0.996±0.4
Symbols32±2.18±0.58±0.58±0.55±1.616±1.595±1.7
synthetic_control95±0.795±0.495±0.495±0.420±495±0.7100±0.4
ToeSegmentation154±1.831±2.531±2.531±2.54±0.739±297±0.7
ToeSegmentation245±5.235±5.935±5.935±5.911±2.541±4.390±2.5
Trace30±2.113±9.713±9.713±9.72±1.637±8.698±0
TwoLeadECG8±4.82±0.62±0.62±0.61±0.54±1.7100±0.3
Two_Patterns68±1.942±6.242±6.242±6.26±1.156±3.896±1
wafer17±11.87±7.87±7.87±7.82±0.211±10.8100±0.1
Wine34±1625±8.425±8.425±8.425±8.425±8.476±8.4
WordsSynonyms15±3.114±114±114±116±0.414±1.454±1.3
Worms26±221±1.521±1.521±1.519±0.925±0.463±2
WormsTwoClass54±2.729±229±229±227±232±1.475±1.4