/
generateFigures.py
204 lines (167 loc) · 10.6 KB
/
generateFigures.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
import os
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
import seaborn as sns
import pdb
plt.rcParams.update({'font.size': 14})
figureDirectory = "./Figures/"
excelFile = "experimentSummary.xlsx"
dataFrame = pd.read_excel(excelFile)
if not os.path.isdir(figureDirectory):
os.mkdir(figureDirectory)
architecture = ["DenseNet121", "DPN92", "GoogLeNet", "MobileNetV2", "ResNet18", "ResNet50", "ResNeXt29_2x64d", "SENet18"]
replicateImbalance = [True, False]
classBalance = [2, 5, 7, 10, 12, 15, 20, 25, 35, 50]
KValues = [2, 5, 10, 20, 50, 75, 90, 100, 110, 125, 200, 400]
fixedClassBalance = 50
fixedK = 100
for rowTable in dataFrame.iterrows():
CB = rowTable[1]["Class Balance"]
targetWeight = CB.replace("[50, 50, 50, 50, 50, 50, ", "").replace(", 50, 50, 50]", "")
dataFrame.at[rowTable[0], "Class Balance"] = int(targetWeight)
#Q1: (Fixed Class Balance) Is the effectiveness of KNN Defense Model Specific?
Q1_dataFrame = dataFrame[(dataFrame["Class Balance"] == fixedClassBalance)][["Model Architecture", "K Value", "Poison Success on Target Image"]]
Q1_Statistics = {"Model Architecture" : [],
"Normalized-K Ratio" : [],
"Attack Success Rate" : []}
for modelName in architecture:
for K in KValues:
Q1_Statistics["Model Architecture"].append(modelName)
Q1_Statistics["Normalized-K Ratio"].append((K / fixedClassBalance))
tempDataFrame = Q1_dataFrame[(Q1_dataFrame["Model Architecture"] == modelName) & (Q1_dataFrame["K Value"] == K)]
Q1_Statistics["Attack Success Rate"].append(np.sum(tempDataFrame["Poison Success on Target Image"] == True) / float(tempDataFrame.shape[0]))
Q1_Plot = pd.DataFrame.from_dict(Q1_Statistics)
ax = sns.pointplot(x="Normalized-K Ratio", y="Attack Success Rate", hue="Model Architecture", data=Q1_Plot)
ax.legend(loc='upper right', ncol=2, fontsize="small")
plt.title("Transfer Convex Polytope Attack Success")
plt.savefig(figureDirectory + "fixedClassBalance_ModelAgnostic.pdf")
plt.clf()
#Q2: (Fixed Class Balance) Matthews Correlation Coefficient of All Models?
Q2_dataFrame = dataFrame[(dataFrame["Class Balance"] == fixedClassBalance)][["K Value", "Model Architecture", "Matthews Correlation Coefficient"]]
Q2_Statistics = {"Model Architecture" : [],
"Normalized-K Ratio" : [],
"Matthews Correlation Coefficient" : []}
for modelName in architecture:
for K in KValues:
Q2_Statistics["Model Architecture"].append(modelName)
Q2_Statistics["Normalized-K Ratio"].append((K / fixedClassBalance))
tempDataFrame = Q2_dataFrame[(Q2_dataFrame["Model Architecture"] == modelName) & (Q2_dataFrame["K Value"] == K)]
Q2_Statistics["Matthews Correlation Coefficient"].append(np.mean(tempDataFrame["Matthews Correlation Coefficient"]))
Q2_Plot = pd.DataFrame(Q2_Statistics)
ax = sns.pointplot(x="Normalized-K Ratio", y="Matthews Correlation Coefficient", hue="Model Architecture", ci=None, data=Q2_Plot)
ax.get_legend().set_visible(False)
plt.title("Deep-KNN Poison Filtering Success")
plt.savefig(figureDirectory + "fixedClassBalance_CorrelationCoefficient.pdf")
plt.clf()
#Q3: (Fixed Class Balance) Test Accuracy of All Models?
Q3_dataFrame = dataFrame[(dataFrame["Class Balance"] == fixedClassBalance)][["K Value", "Model Architecture", "Test Accuracy"]]
Q3_Statistics = {"Model Architecture" : [],
"Normalized-K Ratio" : [],
"Test Accuracy" : []}
for modelName in architecture:
for K in KValues:
Q3_Statistics["Model Architecture"].append(modelName)
Q3_Statistics["Normalized-K Ratio"].append((K / fixedClassBalance))
tempDataFrame = Q3_dataFrame[(Q3_dataFrame["Model Architecture"] == modelName) & (Q3_dataFrame["K Value"] == K)]
Q3_Statistics["Test Accuracy"].append(np.mean(tempDataFrame["Test Accuracy"]))
Q3_Plot = pd.DataFrame(Q3_Statistics)
ax = sns.pointplot(x="Normalized-K Ratio", y="Test Accuracy", hue="Model Architecture", ci=None, data=Q3_Plot)
ax.get_legend().set_visible(False)
plt.title("Deep-KNN Test Accuracy")
plt.savefig(figureDirectory + "fixedClassBalance_TestAccuracy.pdf")
plt.clf()
plt.rcParams.update({'font.size': 20})
#Q4: (Fixed K) Is the effectiveness of KNN Defense Model Specific?
Q4_dataFrame = dataFrame[(dataFrame["K Value"] == fixedK)][["Model Architecture", "Class Balance", "Replicate Imbalance", "Poison Success on Target Image"]]
Q4_Statistics = {"Model Architecture" : [],
"Class Balance Ratio" : [],
"Replicate Imbalance" : [],
"Attack Success Rate" : []}
for modelName in architecture:
for targetWeight in classBalance:
for replicate in replicateImbalance:
Q4_Statistics["Model Architecture"].append(modelName)
Q4_Statistics["Class Balance Ratio"].append(targetWeight / fixedClassBalance)
Q4_Statistics["Replicate Imbalance"].append(replicate)
tempDataFrame = Q4_dataFrame[(Q4_dataFrame["Model Architecture"] == modelName) & (Q4_dataFrame["Replicate Imbalance"] == replicate) & (Q4_dataFrame["Class Balance"] == targetWeight)]
Q4_Statistics["Attack Success Rate"].append(np.sum(tempDataFrame["Poison Success on Target Image"] == True) / float(tempDataFrame.shape[0]))
Q4_Plot = pd.DataFrame.from_dict(Q4_Statistics)
ax = sns.catplot(x="Class Balance Ratio", y="Attack Success Rate", hue="Model Architecture", data=Q4_Plot, col="Replicate Imbalance", order=[i / fixedClassBalance for i in classBalance], kind="point", legend=False)
plt.subplots_adjust(top=0.85)
ax.fig.suptitle("Transfer Convex Polytope Attack Success")
ax.set_xticklabels(rotation=45)
#plt.legend(loc='upper right', ncol=2)
plt.savefig(figureDirectory + "fixedK_modelAgnostic.pdf")
plt.clf()
#Q5: (Fixed K) Matthews Correlation Coefficient of All Models?
Q5_dataFrame = dataFrame[(dataFrame["K Value"] == fixedK)][["Class Balance", "Replicate Imbalance", "Model Architecture", "Matthews Correlation Coefficient"]]
Q5_Statistics = {"Model Architecture" : [],
"Class Balance Ratio" : [],
"Replicate Imbalance" : [],
"Matthews Correlation Coefficient" : []}
for modelName in architecture:
for targetWeight in classBalance:
for replicate in replicateImbalance:
Q5_Statistics["Model Architecture"].append(modelName)
Q5_Statistics["Class Balance Ratio"].append(targetWeight / fixedClassBalance)
Q5_Statistics["Replicate Imbalance"].append(replicate)
tempDataFrame = Q5_dataFrame[(Q5_dataFrame["Model Architecture"] == modelName) & (Q5_dataFrame["Replicate Imbalance"] == replicate) & (Q5_dataFrame["Class Balance"] == targetWeight)]
Q5_Statistics["Matthews Correlation Coefficient"].append(np.mean(tempDataFrame["Matthews Correlation Coefficient"]))
Q5_Plot = pd.DataFrame(Q5_Statistics)
ax = sns.catplot(x="Class Balance Ratio", y="Matthews Correlation Coefficient", hue="Model Architecture", data=Q5_Plot, col="Replicate Imbalance", order=[i / fixedClassBalance for i in classBalance], kind="point", ci=None, legend=False)
plt.subplots_adjust(top=0.85)
ax.fig.suptitle("Deep-KNN Poison Filtering Success")
ax.set_xticklabels(rotation=45)
plt.savefig(figureDirectory + "fixedK_CorrelationCoefficient.pdf")
plt.clf()
#Q6: (Fixed K Value) Test Accuracy of All Models?
Q6_dataFrame = dataFrame[(dataFrame["K Value"] == fixedK)][["Class Balance", "Replicate Imbalance", "Model Architecture", "Test Accuracy"]]
Q6_Statistics = {"Model Architecture" : [],
"Class Balance Ratio" : [],
"Replicate Imbalance" : [],
"Test Accuracy" : []}
for modelName in architecture:
for targetWeight in classBalance:
for replicate in replicateImbalance:
Q6_Statistics["Model Architecture"].append(modelName)
Q6_Statistics["Class Balance Ratio"].append(targetWeight / fixedClassBalance)
Q6_Statistics["Replicate Imbalance"].append(replicate)
tempDataFrame = Q6_dataFrame[(Q6_dataFrame["Model Architecture"] == modelName) & (Q6_dataFrame["Replicate Imbalance"] == replicate) & (Q6_dataFrame["Class Balance"] == targetWeight)]
Q6_Statistics["Test Accuracy"].append(np.mean(tempDataFrame["Test Accuracy"]))
Q6_Plot = pd.DataFrame(Q6_Statistics)
ax = sns.catplot(x="Class Balance Ratio", y="Test Accuracy", hue="Model Architecture", data=Q6_Plot, col="Replicate Imbalance", order=[i / fixedClassBalance for i in classBalance], kind="point", ci=None, legend=False)
plt.subplots_adjust(top=0.85)
ax.fig.suptitle("Deep-KNN Test Accuracy")
ax.set_xticklabels(rotation=45)
plt.savefig(figureDirectory + "fixedK_TestAccuracy.pdf")
plt.clf()
plt.rcParams.update({'font.size': 13})
dataFrame = pd.read_csv("defenseRates.csv").rename(columns={"Network":"Model Architecture", "Defense":"Defense",
"Attack success rate":"Attack Success Rate",
"Defense_rate":"Defense Success Rate"})
for rowTable in dataFrame.iterrows():
if "k-NN" in rowTable[1]["Defense"]:
dataFrame.at[rowTable[0], "Defense"] = "Deep K-NN"
elif "SVM" in rowTable[1]["Defense"]:
dataFrame.at[rowTable[0], "Defense"] = "One Class SVM"
elif "L2" in rowTable[1]["Defense"]:
dataFrame.at[rowTable[0], "Defense"] = "L2-Norm Outliers"
elif "Adversarial" in rowTable[1]["Defense"]:
dataFrame.at[rowTable[0], "Defense"] = "Adversarial Training"
elif "Random" in rowTable[1]["Defense"]:
dataFrame.at[rowTable[0], "Defense"] = "Random Point Eviction"
elif "No" in rowTable[1]["Defense"]:
dataFrame.at[rowTable[0], "Defense"] = "No Defense"
if "ResNeXt" in rowTable[1]["Model Architecture"]:
dataFrame.at[rowTable[0], "Model Architecture"] = "ResNeXt"
if "DenseNet" in rowTable[1]["Model Architecture"]:
dataFrame.at[rowTable[0], "Model Architecture"] = "DenseNet"
if "MobileNet" in rowTable[1]["Model Architecture"]:
dataFrame.at[rowTable[0], "Model Architecture"] = "MobileNet"
if "SENet" in rowTable[1]["Model Architecture"]:
dataFrame.at[rowTable[0], "Model Architecture"] = "SENet"
ax = sns.barplot(x="Model Architecture", y="Defense Success Rate", hue="Defense", data=dataFrame)
ax.legend(loc='upper center', bbox_to_anchor=(0.5, 1.10), ncol=3, fontsize="x-small")
ax.set_xticklabels(ax.get_xticklabels(), rotation=17)
plt.savefig(figureDirectory + "ConvexPolytope_CompareDefense.pdf")