-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
335 lines (302 loc) · 14.5 KB
/
index.html
File metadata and controls
335 lines (302 loc) · 14.5 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
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description" content="Pushing the Frontier of Black-Box LVLM Attacks via Fine-Grained Detail Targeting">
<meta name="keywords" content="LVLM, adversarial attack, black-box attack, vision-language model">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>M-Attack V2: Pushing the Frontier of Black-Box LVLM Attacks via Fine-Grained Detail Targeting</title>
<!-- Global site tag (gtag.js) - Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-PYVRSFMDRL"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() {
dataLayer.push(arguments);
}
gtag('js', new Date());
gtag('config', 'G-PYVRSFMDRL');
</script>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="stylesheet" href="./static/css/bulma.min.css">
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
<script src="./static/js/index.js"></script>
<script>
MathJax = {
tex: { inlineMath: [['$', '$'], ['\\(', '\\)']] },
svg: { fontCache: 'global' }
};
</script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-svg.js"></script>
</head>
<body>
<nav class="navbar" role="navigation" aria-label="main navigation">
<div class="navbar-brand">
<a role="button" class="navbar-burger" aria-label="menu" aria-expanded="false">
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
</a>
</div>
<div class="navbar-menu">
<div class="navbar-start" style="flex-grow: 1; justify-content: center;">
<a class="navbar-item" href="https://github.com/VILA-Lab">
<span class="icon">
<i class="fas fa-home"></i>
</span>
</a>
</div>
</div>
</nav>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title">Pushing the Frontier of Black-Box LVLM Attacks via Fine-Grained
Detail Targeting</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://scholar.google.com/citations?user=PliLuD4AAAAJ&hl=en">Xiaohan Zhao</a>,</span>
<span class="author-block">
<a href="https://scholar.google.com/citations?user=J-VsziYAAAAJ&hl=en">Zhaoyi Li</a>,</span>
<span class="author-block">
<a href="https://scholar.google.com/citations?user=tEaSCzYAAAAJ&hl=en">Yaxin Luo</a>,
</span>
<span class="author-block">
<a href="https://jiachengcui.com/">Jiacheng Cui</a>,
</span>
<span class="author-block">
<a href="https://zhiqiangshen.com">Zhiqiang Shen</a><sup>†</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block">VILA Lab, Department of Machine Learning, MBZUAI</span>
</div>
<div class="is-size-6 publication-authors">
<span class="author-block"><sup>†</sup>Corresponding Author</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="./static/paper.pdf" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Paper</span>
</a>
</span>
<span class="link-block">
<a href="#" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/VILA-Lab/M-Attack-V2"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Teaser: Improvement over M-Attack -->
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<img src="./static/images/improvement.png" alt="M-Attack-V2 improvement over M-Attack" width="100%">
<h2 class="subtitle has-text-centered">
<b>M-Attack-V2</b> significantly improves attack success rate (ASR) and keyword matching rate (KMR)
over M-Attack across state-of-the-art commercial black-box models including Claude 4, Gemini 2.5, and GPT-5.
</h2>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Unexpectedly Low Gradient Similarity -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Unexpectedly Low Gradient Similarity</h2>
<div class="content has-text-justified">
<p>
Local-level matching methods exhibit <b>near-zero gradient cosine similarity</b> between iterations,
even with significant spatial overlap between crops. This stems from ViTs' translation sensitivity
and an overlooked <b>asymmetry</b>: source crops reshape the pixel-space gradient landscape,
while target crops merely shift the feature-space reference.
</p>
</div>
<img src="./static/images/similarity.png" alt="Gradient cosine similarity analysis." style="width: 100%;" />
<p class="has-text-centered is-size-6" style="margin-top: 0.5rem; color: #666;">
<i>(a) Gradient similarity vs. IoU between two crops. (b) Cosine similarity of consecutive source gradients
across iterations.</i>
</p>
</div>
</div>
<!-- Asymmetric Matching Formulation -->
<div class="columns is-centered has-text-centered" style="margin-top: 2rem;">
<div class="column is-four-fifths">
<h2 class="title is-3">Asymmetric Matching over Expectation</h2>
<div class="content has-text-justified">
<p>
We reformulate the objective as an expectation over local transformations within an asymmetric framework:
</p>
<p style="text-align: center;">
$$\min_{\lVert \mathbf{X}_\text{sou} \rVert_p \le \epsilon} \mathbb{E}_{\mathcal{T} \sim \mathcal{D},\, y
\sim \mathcal{Y}} \left[ \mathcal{L}\!\left(f\!\left(\mathcal{T}(\mathbf{X}_{\text{sou}})\right),\,
y\right) \right]$$
</p>
<p>
where $\mathcal{D}$ is the distribution of local transformations and $\mathcal{Y}$ the target semantic
distribution. This highlights the intrinsic asymmetry: embedding content $y$ into a locally transformed
source $\mathcal{T}(\mathbf{X}_{\text{sou}})$. Our two enhancements, <i>Multi-Crop Alignment</i> (MCA) and
<i>Auxiliary Target Alignment</i> (ATA), improve the expectation estimation and the sampling quality
of $\mathcal{Y}$, respectively.
</p>
</div>
</div>
</div>
<!-- Gradient Denoising via Multi-Crop Alignment -->
<div class="columns is-centered has-text-centered" style="margin-top: 2rem;">
<div class="column is-four-fifths">
<h2 class="title is-3">Gradient Denoising via Multi-Crop Alignment</h2>
<div class="content has-text-justified">
<p>
<b>MCA</b> averages gradients from $K$ independent crops per iteration, yielding a low-variance
estimate of the expected gradient. This produces smoother gradient patterns and accelerates convergence
compared to single-crop alignment.
</p>
</div>
<img src="./static/images/mca.png" alt="Multi-Crop Alignment comparison." style="width: 100%;" />
<p class="has-text-centered is-size-6" style="margin-top: 0.5rem; color: #666;">
<i>(a) Optimization trajectories with different K. (b) Gradient patterns: single-crop (M-Attack) vs.
multi-crop (M-Attack-V2).</i>
</p>
</div>
</div>
<!-- ATA + Algorithm: two-column layout -->
<div class="columns is-centered" style="margin-top: 2rem;">
<div class="column is-three-fifths">
<h2 class="title is-3 has-text-centered">Auxiliary Target Alignment</h2>
<div class="content has-text-justified">
<p>
Selecting a representative target embedding $y \in \mathcal{Y}$ is challenging since $\mathcal{Y}$
is unobservable. M-Attack explores via transformed views of the target, but radical crops drift too far
while conservative ones provide little signal.
</p>
<p>
<b>ATA</b> introduces $P$ auxiliary images $\{\mathbf{X}_\text{aux}^{(p)}\}_{p=1}^P$ as additional
semantic anchors. With mild transformations $\tilde{\mathcal{T}} \sim \tilde{\mathcal{D}}$ applied to
each anchor, the combined objective becomes:
</p>
<p style="text-align: center;">
$$\hat{\mathcal{L}} = \frac{1}{K} \sum_{k=1}^{K} \Big[
\mathcal{L}(f(\mathcal{T}_k(\mathbf{X}_\text{sou})), y_0)
+ \frac{\lambda}{P} \sum_{p=1}^{P} \mathcal{L}(f(\mathcal{T}_k(\mathbf{X}_{\text{sou}})), \tilde{y}_p)
\Big]$$
</p>
<p>
where $y_0 = f(\hat{\mathcal{T}}_0(\mathbf{X}_\text{tar}))$, $\tilde{y}_p =
f(\tilde{\mathcal{T}}_p(\mathbf{X}_\text{aux}^{(p)}))$,
and $\lambda \in [0,1]$ interpolates between target fidelity and auxiliary diversity.
ATA achieves a better exploration-exploitation balance by allocating its shift budget toward
semantically meaningful exploration via the auxiliary set.
</p>
</div>
</div>
<div class="column is-two-fifths has-text-centered"
style="display: flex; align-items: center; justify-content: center;">
<img src="./static/images/algorithim.png" alt="Algorithm: M-Attack-V2 pipeline."
style="width: 100%; max-width: 360px;" />
</div>
</div>
<!-- Experimental Results -->
<div class="columns is-centered has-text-centered" style="margin-top: 2rem;">
<div class="column">
<h2 class="title is-3">Experimental Results</h2>
<div class="content has-text-justified">
<p>
M-Attack-V2 consistently outperforms all existing methods across GPT-5, Claude 4.0-thinking,
and Gemini 2.5-Pro, achieving the highest attack success rates with strong imperceptibility.
</p>
</div>
<img src="./static/images/main_result.png" alt="Main results comparison table." style="width: 100%;" />
<p class="has-text-centered is-size-6" style="margin-top: 0.5rem; color: #666;">
<i>Comparison with state-of-the-art approaches on commercial black-box LVLMs.</i>
</p>
</div>
</div>
<!-- Visualization -->
<div class="columns is-centered has-text-centered" style="margin-top: 2rem;">
<div class="column is-four-fifths">
<h2 class="title is-3">Visualization of Adversarial Samples</h2>
<div class="content has-text-justified">
<p>
Visual comparison across methods. M-Attack-V2 produces more effective yet more imperceptible
perturbations.
</p>
</div>
<img src="./static/images/adv_samples.png" alt="Adversarial sample visualization." style="width: 100%;" />
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@article{zhao2026pushingfrontierblackboxlvlm,
title={Pushing the Frontier of Black-Box LVLM Attacks via Fine-Grained Detail Targeting},
author={Zhao, Xiaohan and Li, Zhaoyi and Luo, Yaxin and Cui, Jiacheng and Shen, Zhiqiang},
journal={arXiv preprint arXiv:2602.17645},
year={2026}
}</code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="content has-text-centered">
<a class="icon-link" href="https://github.com/VILA-Lab" class="external-link" disabled>
<i class="fab fa-github"></i>
</a>
</div>
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This website is licensed under a <a rel="license"
href="http://creativecommons.org/licenses/by-sa/4.0/">Creative
Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
<p>
This means you are free to borrow the <a href="https://github.com/nerfies/nerfies.github.io">source
code</a> of this website,
we just ask that you link back to this page in the footer.
Please remember to remove the analytics code included in the header of the website which
you do not want on your website.
</p>
</div>
</div>
</div>
</div>
</footer>
</body>
</html>