-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
182 lines (175 loc) · 7.37 KB
/
index.html
File metadata and controls
182 lines (175 loc) · 7.37 KB
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
---
layout: default
nav_id: home
---
<div id="page-root"></div>
<script type="module" src="{{ "/assets/js/page-home.js" | relative_url }}"></script>
---
layout: default
nav_id: home
---
<div id="page-root"></div>
<script type="module" src="{{ "/assets/js/page-home.js" | relative_url }}"></script>
layout: default
---
<section id="home" class="agb-hero">
<div class="container">
<p class="agb-kicker">Research Project Website</p>
<h1>Adversarial Graph Benchmark (AGB)</h1>
<p class="agb-subtitle">Evaluating GNN Robustness via Topology-Aware Adversarial Attacks</p>
<p class="agb-lead">
AGB presents GOttack, a universal adversarial framework for graph neural networks that leverages orbit-aware
topology signals to discover high-impact perturbation candidates. The benchmark studies how graph structure governs
robustness under constrained edge manipulations.
</p>
<a class="btn btn-primary btn-lg" href="#method">Explore Method</a>
</div>
</section>
<section class="agb-section" id="motivation">
<div class="container">
<h2>Motivation</h2>
<div class="row">
<div class="col-md-4">
<div class="agb-card">
<h3>Graph Neural Networks</h3>
<p>GNNs aggregate node attributes and neighborhood structure to perform relational tasks such as node classification.</p>
</div>
</div>
<div class="col-md-4">
<div class="agb-card">
<h3>Adversarial Vulnerability</h3>
<p>Even sparse edge perturbations can alter message passing trajectories and produce severe prediction drift.</p>
</div>
</div>
<div class="col-md-4">
<div class="agb-card">
<h3>Role of Topology</h3>
<p>Structural roles determine attack sensitivity, making topology-aware perturbation design essential for robust evaluation.</p>
</div>
</div>
</div>
</div>
</section>
<section class="agb-section agb-section-alt" id="method">
<div class="container">
<h2>Methodology</h2>
<div class="agb-panel">
<p><strong>Graph definition.</strong> We model each graph as <code>G = (V, E, X)</code>, with nodes <code>V</code>, edges <code>E</code>, and node features <code>X</code>.</p>
<p><strong>Node classification setup.</strong> An attacker modifies graph structure through limited edge insertions and deletions to induce target misclassification.</p>
<p><strong>Structural attack context.</strong> The candidate search is combinatorial; prioritization in topological space is the key design decision.</p>
</div>
<div class="row">
<div class="col-md-7">
<div class="agb-card">
<h3>GOttack</h3>
<ul>
<li>Orbit-based node grouping captures local structural roles.</li>
<li>A 73-dimensional Graph Orbit Vector enables topology-aware ranking.</li>
<li>Orbit 15 and orbit 18 are emphasized as high-impact regions.</li>
<li>Periphery nodes are shown to drive substantial misclassification.</li>
</ul>
</div>
</div>
<div class="col-md-5">
<div class="agb-card">
<h3>Core Hypothesis</h3>
<p>Orbit-informed perturbation selection improves universal attack strength while preserving sparse, efficient edits.</p>
</div>
</div>
</div>
</div>
</section>
<section class="agb-section" id="pipeline">
<div class="container">
<h2>Pipeline</h2>
<div class="agb-flow">
<div class="agb-flow-step">Graph</div>
<div class="agb-flow-arrow">→</div>
<div class="agb-flow-step">Orbit Detection</div>
<div class="agb-flow-arrow">→</div>
<div class="agb-flow-step">Candidate Selection</div>
<div class="agb-flow-arrow">→</div>
<div class="agb-flow-step">Edge Perturbation</div>
<div class="agb-flow-arrow">→</div>
<div class="agb-flow-step">Misclassification</div>
</div>
</div>
</section>
<section class="agb-section agb-section-alt" id="results">
<div class="container">
<h2>Results</h2>
<div class="row">
<div class="col-md-4">
<div class="agb-card">
<h3>Attack Strength</h3>
<p>GOttack achieves higher misclassification performance across evaluated tasks.</p>
</div>
</div>
<div class="col-md-4">
<div class="agb-card">
<h3>Model Coverage</h3>
<p>The method remains effective across GCN, GIN, and GraphSAGE.</p>
</div>
</div>
<div class="col-md-4">
<div class="agb-card">
<h3>Efficiency</h3>
<p>Runtime is approximately 85% of competing attack methods while preserving stronger outcomes.</p>
</div>
</div>
</div>
<div class="agb-panel">
<h3>Evaluation Datasets</h3>
<div class="agb-tag-list">
<span class="agb-tag">Cora</span>
<span class="agb-tag">Citeseer</span>
<span class="agb-tag">Pubmed</span>
<span class="agb-tag">BlogCatalog</span>
<span class="agb-tag">Polblogs</span>
</div>
</div>
</div>
</section>
<section class="agb-section" id="comparison">
<div class="container">
<h2>Comparison</h2>
<div class="row">
<div class="col-sm-3">
<div class="agb-card">
<h3>Nettack</h3>
<p>Strong baseline with targeted perturbations, but limited topology-role awareness.</p>
</div>
</div>
<div class="col-sm-3">
<div class="agb-card">
<h3>FGA</h3>
<p>Gradient-guided adversarial edits with effective local search.</p>
</div>
</div>
<div class="col-sm-3">
<div class="agb-card">
<h3>SGA</h3>
<p>Scalable perturbation strategy balancing cost and attack utility.</p>
</div>
</div>
<div class="col-sm-3">
<div class="agb-card agb-card-highlight">
<h3>GOttack</h3>
<p>Best overall attack quality and efficiency through orbit-aware selection.</p>
</div>
</div>
</div>
</div>
</section>
<section class="agb-section agb-section-alt" id="insights">
<div class="container">
<h2>Insights</h2>
<div class="agb-panel">
<ul>
<li>Orbit-based attacks are highly effective under constrained perturbation budgets.</li>
<li>Orbits 15 and 18 identify structurally critical nodes for attack planning.</li>
<li>Graph topology is a major factor in practical GNN robustness.</li>
</ul>
</div>
</div>
</section>