def analyze_clusters(self, **kwargs):
"""
This function analyzes the clusters according to the provided method
:param kwargs: a dictionary of cluster-analysis-specific parameters
:type kwargs: `dict`
:return: assigned_clean_by_class, an array of arrays that contains what data points where classified as clean.
:rtype: `np.ndarray`
"""
self.set_params(**kwargs)
if not self.clusters_by_class:
self.cluster_activations()
analyzer = ClusteringAnalyzer()
if self.cluster_analysis == 'smaller':
self.assigned_clean_by_class, self.poisonous_clusters \
= analyzer.analyze_by_size(self.clusters_by_class)
elif self.cluster_analysis == 'relative-size':
self.assigned_clean_by_class, self.poisonous_clusters \
= analyzer.analyze_by_relative_size(self.clusters_by_class)
elif self.cluster_analysis == 'distance':
self.assigned_clean_by_class, self.poisonous_clusters \
= analyzer.analyze_by_distance(self.clusters_by_class,
separated_activations=self.red_activations_by_class)
elif self.cluster_analysis == 'silhouette-scores':
self.assigned_clean_by_class, self.poisonous_clusters \
= analyzer.analyze_by_sihouette_score(self.clusters_by_class,
reduced_activations_by_class=self.red_activations_by_class)
else:
raise ValueError(
"Unsupported cluster analysis technique " + self.cluster_analysis)
return self.assigned_clean_by_class
def test_relative_size_analyzer(self):
nb_clusters = 2
nb_classes = 4
clusters_by_class = [[[] for x in range(nb_clusters)]
for y in range(nb_classes)]
clusters_by_class[0] = [0, 1, 1, 1, 1] # Class 0
clusters_by_class[1] = [1, 0, 0, 0, 0] # Class 1
clusters_by_class[2] = [0, 0, 0, 0, 1] # Class 2
clusters_by_class[3] = [0, 0, 1, 1, 1] # Class 3
analyzer = ClusteringAnalyzer()
assigned_clean_by_class, poison_clusters = analyzer.analyze_by_relative_size(
clusters_by_class)
# print("clusters_by_class")
# print(clusters_by_class)
# print("assigned_clean_by_class")
# print(assigned_clean_by_class)
# print("poison_clusters")
# print(poison_clusters)
clean = 0
poison = 1
# For class 0, cluster 0 should be marked as poison.
self.assertEqual(poison_clusters[0][0], poison)
# For class 0, cluster 1 should be marked as clean.
self.assertEqual(poison_clusters[0][1], clean)
# Inverse relations for class 1
self.assertEqual(poison_clusters[1][0], clean)
self.assertEqual(poison_clusters[1][1], poison)
self.assertEqual(poison_clusters[2][0], clean)
self.assertEqual(poison_clusters[2][1], poison)
self.assertEqual(poison_clusters[3][0], clean)
self.assertEqual(poison_clusters[3][1], clean)
poison = 0
self.assertEqual(assigned_clean_by_class[0][0], poison)
self.assertEqual(assigned_clean_by_class[1][0], poison)
self.assertEqual(assigned_clean_by_class[2][4], poison)
self.assertEqual(sum(assigned_clean_by_class[3]),
len(assigned_clean_by_class[3]))
def analyze_clusters(self, **kwargs):
"""
This function analyzes the clusters according to the provided method
:param kwargs: a dictionary of cluster-analysis-specific parameters
:type kwargs: `dict`
:return: (report, assigned_clean_by_class), where the report is a json object and assigned_clean_by_class
is an array of arrays that contains what data points where classified as clean.
:rtype: `tuple(json, np.ndarray)`
"""
self.set_params(**kwargs)
if not self.clusters_by_class:
self.cluster_activations()
analyzer = ClusteringAnalyzer()
if self.cluster_analysis == 'smaller':
self.assigned_clean_by_class, self.poisonous_clusters, report \
= analyzer.analyze_by_size(self.clusters_by_class)
elif self.cluster_analysis == 'relative-size':
self.assigned_clean_by_class, self.poisonous_clusters, report \
= analyzer.analyze_by_relative_size(self.clusters_by_class)
elif self.cluster_analysis == 'distance':
self.assigned_clean_by_class, self.poisonous_clusters, report \
= analyzer.analyze_by_distance(self.clusters_by_class,
separated_activations=self.red_activations_by_class)
elif self.cluster_analysis == 'silhouette-scores':
self.assigned_clean_by_class, self.poisonous_clusters, report \
= analyzer.analyze_by_silhouette_score(self.clusters_by_class,
reduced_activations_by_class=self.red_activations_by_class)
else:
raise ValueError(
"Unsupported cluster analysis technique " + self.cluster_analysis)
# Add to the report current parameters used to run the defence and the analysis summary
report = dict(list(report.items()) + list(self.get_params().items()))
import json
jreport = json.dumps(report)
return jreport, self.assigned_clean_by_class
def test_relative_size_analyzer(self):
nb_clusters = 2
nb_classes = 4
clusters_by_class = [[[] for x in range(nb_clusters)]
for y in range(nb_classes)]
clusters_by_class[0] = [0, 1, 1, 1, 1] # Class 0
clusters_by_class[1] = [1, 0, 0, 0, 0] # Class 1
clusters_by_class[2] = [0, 0, 0, 0, 1] # Class 2
clusters_by_class[3] = [0, 0, 1, 1, 1] # Class 3
analyzer = ClusteringAnalyzer()
assigned_clean_by_class, poison_clusters, report = analyzer.analyze_by_relative_size(
clusters_by_class)
# print("clusters_by_class")
# print(clusters_by_class)
# print("assigned_clean_by_class")
# print(assigned_clean_by_class)
# print("poison_clusters")
# print(poison_clusters)
clean = 0
poison = 1
# For class 0, cluster 0 should be marked as poison.
self.assertEqual(poison_clusters[0][0], poison)
# For class 0, cluster 1 should be marked as clean.
self.assertEqual(poison_clusters[0][1], clean)
self.assertEqual(report['Class_0']['cluster_0']['suspicious_cluster'],
True)
self.assertEqual(report['Class_0']['cluster_1']['suspicious_cluster'],
False)
total = len(clusters_by_class[0])
c1 = sum(clusters_by_class[0])
self.assertEqual(report['Class_0']['cluster_0']['ptc_data_in_cluster'],
round((total - c1) / total, 2))
self.assertEqual(report['Class_0']['cluster_1']['ptc_data_in_cluster'],
round(c1 / total, 2))
# Inverse relations for class 1
self.assertEqual(poison_clusters[1][0], clean)
self.assertEqual(poison_clusters[1][1], poison)
self.assertEqual(report['Class_1']['cluster_0']['suspicious_cluster'],
False)
self.assertEqual(report['Class_1']['cluster_1']['suspicious_cluster'],
True)
total = len(clusters_by_class[1])
c1 = sum(clusters_by_class[1])
self.assertEqual(report['Class_1']['cluster_0']['ptc_data_in_cluster'],
round((total - c1) / total, 2))
self.assertEqual(report['Class_1']['cluster_1']['ptc_data_in_cluster'],
round(c1 / total, 2))
self.assertEqual(poison_clusters[2][0], clean)
self.assertEqual(poison_clusters[2][1], poison)
self.assertEqual(report['Class_2']['cluster_0']['suspicious_cluster'],
False)
self.assertEqual(report['Class_2']['cluster_1']['suspicious_cluster'],
True)
total = len(clusters_by_class[2])
c1 = sum(clusters_by_class[2])
self.assertEqual(report['Class_2']['cluster_0']['ptc_data_in_cluster'],
round((total - c1) / total, 2))
self.assertEqual(report['Class_2']['cluster_1']['ptc_data_in_cluster'],
round(c1 / total, 2))
self.assertEqual(poison_clusters[3][0], clean)
self.assertEqual(poison_clusters[3][1], clean)
self.assertEqual(report['Class_3']['cluster_0']['suspicious_cluster'],
False)
self.assertEqual(report['Class_3']['cluster_1']['suspicious_cluster'],
False)
total = len(clusters_by_class[3])
c1 = sum(clusters_by_class[3])
self.assertEqual(report['Class_3']['cluster_0']['ptc_data_in_cluster'],
round((total - c1) / total, 2))
self.assertEqual(report['Class_3']['cluster_1']['ptc_data_in_cluster'],
round(c1 / total, 2))
poison = 0
self.assertEqual(assigned_clean_by_class[0][0], poison)
self.assertEqual(assigned_clean_by_class[1][0], poison)
self.assertEqual(assigned_clean_by_class[2][4], poison)
self.assertEqual(sum(assigned_clean_by_class[3]),
len(assigned_clean_by_class[3]))