-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
577 lines (548 loc) · 31.5 KB
/
Copy pathindex.html
File metadata and controls
577 lines (548 loc) · 31.5 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
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>SCAM: A Real-World Typographic Robustness Evaluation for Multimodal Foundation Models</title>
<link rel="stylesheet" href="styles.css">
</head>
<body>
<header class="header">
<div class="container">
<h1>SCAM: A Real-World Typographic Robustness Evaluation for Multimodal Foundation Models</h1>
<div class="authors">
<p class="header-label">Authors</p>
<p>
<a href="https://www.linkedin.com/in/justus-westerhoff/">Justus Westerhoff</a><sup>1,2</sup>,
<a href="https://github.com/blinapurellku">Erblina Purelku</a><sup>1,3,5</sup>,
<a href="https://github.com/jakhac">Jakob Hackstein</a><sup>1,3</sup>,
<a href="https://github.com/JonasLoos/">Jonas Loos</a><sup>1,3</sup>,
<a href="https://www.linkedin.com/in/leo-pinetzki/">Leo Pinetzki</a><sup>1,3</sup>,
<a href="https://www.linkedin.com/in/erik-rodner-6914a263">Erik Rodner</a><sup>4</sup>,
<a href="https://www.linkedin.com/in/lorenz-hufe">Lorenz Hufe</a><sup>1,5</sup>
</p>
<p class="header-label">Further Contributors</p>
<p>
<a href="https://www.linkedin.com/in/joseph-tsch%C3%B6rner-12818427a/">Joseph
Tschörner</a><sup>1</sup>,
<a href="https://www.linkedin.com/in/zeynep-altayli-4213a6279/">Zeynep Altayli</a><sup>1</sup>,
<a href="#">Philippa Ringe</a><sup>1</sup>
</p>
<p class="header-label">Affiliations</p>
<p class="affiliations">
<sup>1</sup><a href="https://bliss.berlin/">BLISS e.V.</a>,
<sup>2</sup><a
href="https://projekt.bht-berlin.de/data-science/graduate-school/justus-westerhoff">BHT
Berlin</a>,
<sup>3</sup><a href="https://www.tu.berlin/">TU Berlin</a>,
<sup>4</sup><a href="https://kiwerkstatt.f2.htw-berlin.de/">HTW Berlin</a>,
<sup>5</sup><a href="https://www.hhi.fraunhofer.de/">Fraunhofer HHI</a>
</p>
</div>
<div class="links">
<a href="https://arxiv.org/abs/2504.04893" class="button">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"
style="vertical-align: text-bottom; margin-right: 5px;">
<path d="M14 2H6a2 2 0 0 0-2 2v16a2 2 0 0 0 2 2h12a2 2 0 0 0 2-2V8z"></path>
<polyline points="14 2 14 8 20 8"></polyline>
</svg>
Preprint (arXiv)
</a>
<a href="https://github.com/Bliss-e-V/SCAM" class="button">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"
style="vertical-align: text-bottom; margin-right: 5px;">
<path
d="M9 19c-5 1.5-5-2.5-7-3m14 6v-3.87a3.37 3.37 0 0 0-.94-2.61c3.14-.35 6.44-1.54 6.44-7A5.44 5.44 0 0 0 20 4.77 5.07 5.07 0 0 0 19.91 1S18.73.65 16 2.48a13.38 13.38 0 0 0-7 0C6.27.65 5.09 1 5.09 1A5.07 5.07 0 0 0 5 4.77a5.44 5.44 0 0 0-1.5 3.78c0 5.42 3.3 6.61 6.44 7A3.37 3.37 0 0 0 9 18.13V22">
</path>
</svg>
Code (GitHub)
</a>
<a href="https://huggingface.co/datasets/BLISS-e-V/SCAM" class="button">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"
style="vertical-align: text-bottom; margin-right: 5px;">
<ellipse cx="12" cy="5" rx="9" ry="3"></ellipse>
<path d="M21 12c0 1.66-4 3-9 3s-9-1.34-9-3"></path>
<path d="M3 5v14c0 1.66 4 3 9 3s9-1.34 9-3V5"></path>
</svg>
Dataset (Hugging Face)
</a>
</div>
<br>
<div class="links">
<a href="https://sites.google.com/view/eval-fomo-2-cvpr/accepted-papers" class="button accepted-button">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"
style="vertical-align: text-bottom; margin-right: 5px;">
<path d="M20 6L9 17l-5-5" />
</svg>
Accepted at CVPR 2025 EVAL-FoMo 2
</a>
<a href="https://openreview.net/forum?id=zCcJSErVHH" class="button accepted-button">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"
style="vertical-align: text-bottom; margin-right: 5px;">
<path d="M20 6L9 17l-5-5" />
</svg>
Published at DMLR (OpenReview)
</a>
<a href="https://docs.google.com/forms/d/e/1FAIpQLSe97pPrUUgcRzhnJhmsRzz_s7z3EEDjHK3-MSsqea-LLo2F9A/viewform"
target="_blank" class="button feedback-button">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"
style="vertical-align: text-bottom; margin-right: 5px;">
<path d="M12 20V10"></path>
<path d="M18 20V4"></path>
<path d="M6 20v-16"></path>
</svg>
Feedback
</a>
</div>
</div>
</header>
<section class="interactive-examples">
<div class="container">
<p class="teaser-text"><i>We introduce SCAM datasets to study and evaluate the robustness of multimodal
foundation models against typographic attacks.</i></p>
<div class="explorer-container">
<!-- Left column: Dataset images -->
<div class="dataset-images">
<div class="dataset-variant">
<div class="variant-header-container">
<h3>SCAM</h3>
<p>Real-world Typographic Attack</p>
</div>
<div class="image-container">
<img id="scam-image" src="" alt="SCAM dataset example">
</div>
</div>
<div class="dataset-variant">
<div class="variant-header-container">
<h3>SynthSCAM</h3>
<p>Synthetic Typographic Attack</p>
</div>
<div class="image-container">
<img id="synthscam-image" src="" alt="SynthSCAM dataset example">
</div>
</div>
<div class="dataset-variant">
<div class="variant-header-container">
<h3>NoSCAM</h3>
<p>Removed Attack</p>
</div>
<div class="image-container">
<img id="noscam-image" src="" alt="NoSCAM dataset example">
</div>
</div>
</div>
<!-- Flow arrows from left to center -->
<div class="flow-arrows left-arrows">
<svg class="arrow-svg" viewBox="0 0 100 700" preserveAspectRatio="none">
<path class="arrow-path top-arrow"
d="M 0,60 C 30,60 70,70 90,140 L 100,120 M 90,140 L 70,130" />
<path class="arrow-path middle-arrow"
d="M 0,350 C 30,350 70,350 90,350 L 70,330 M 90,350 L 70,370" />
<path class="arrow-path bottom-arrow"
d="M 0,640 C 30,640 70,630 90,560 L 100,580 M 90,560 L 70,570" />
</svg>
</div>
<!-- Center column: Model selector -->
<div class="model-processor">
<h3>Model</h3>
<div class="mode-selector">
<div class="mode-option">
<input type="radio" id="vlm-mode" name="mode" value="vlm" checked>
<label for="vlm-mode">VLM</label>
</div>
<div class="mode-option">
<input type="radio" id="lvlm-mode" name="mode" value="lvlm">
<label for="lvlm-mode">LVLM</label>
</div>
</div>
<select id="model-selector">
<!-- Options will be populated via JavaScript -->
</select>
<div class="model-info">
<div class="model-property">
<span id="model-params">-</span> params
</div>
<div class="model-property">
<span id="model-image-size">-</span> px
</div>
<div class="model-property" id="model-prompt-container">
<button class="prompt-arrow left-arrow" id="prev-prompt-button"
aria-label="Previous prompt">⟨</button>
<span id="model-prompt">-</span>
<button class="prompt-arrow right-arrow" id="next-prompt-button"
aria-label="Next prompt">⟩</button>
</div>
</div>
</div>
<!-- Flow arrows from center to right -->
<div class="flow-arrows right-arrows">
<svg class="arrow-svg" viewBox="0 0 100 700" preserveAspectRatio="none">
<path class="arrow-path top-arrow" d="M 10,140 C 30,60 70,60 100,60 L 80,75 M 100,60 L 80,45" />
<path class="arrow-path middle-arrow"
d="M 10,350 C 30,350 70,350 100,350 L 80,330 M 100,350 L 80,370" />
<path class="arrow-path bottom-arrow"
d="M 10,560 C 30,640 70,640 100,640 L 80,620 M 100,640 L 80,660" />
</svg>
</div>
<!-- Right column: Results -->
<div class="similarity-results">
<div class="results-container">
<div class="result">
<div class="score-container">
<div class="progress-container">
<div class="label-container">
<span><span class="explorer-label-intro">object:</span> <span
id="scam-object-label">-</span></span>
<span id="scam-object-score" class="score-value">-</span>
</div>
<div class="progress-bar">
<div id="scam-object-bar" class="bar-fill"></div>
</div>
</div>
<div class="progress-container">
<div class="label-container">
<span><span class="explorer-label-intro">attack:</span> <span
id="scam-attack-label">-</span></span>
<span id="scam-attack-score" class="score-value">-</span>
</div>
<div class="progress-bar">
<div id="scam-attack-bar" class="bar-fill"></div>
</div>
</div>
</div>
</div>
<div class="result">
<div class="score-container">
<div class="progress-container">
<div class="label-container">
<span><span class="explorer-label-intro">object:</span> <span
id="synthscam-object-label">-</span></span>
<span id="synthscam-object-score" class="score-value">-</span>
</div>
<div class="progress-bar">
<div id="synthscam-object-bar" class="bar-fill"></div>
</div>
</div>
<div class="progress-container">
<div class="label-container">
<span><span class="explorer-label-intro">attack:</span> <span
id="synthscam-attack-label">-</span></span>
<span id="synthscam-attack-score" class="score-value">-</span>
</div>
<div class="progress-bar">
<div id="synthscam-attack-bar" class="bar-fill"></div>
</div>
</div>
</div>
</div>
<div class="result">
<div class="score-container">
<div class="progress-container">
<div class="label-container">
<span><span class="explorer-label-intro">object:</span> <span
id="noscam-object-label">-</span></span>
<span id="noscam-object-score" class="score-value">-</span>
</div>
<div class="progress-bar">
<div id="noscam-object-bar" class="bar-fill"></div>
</div>
</div>
<div class="progress-container">
<div class="label-container">
<span><span class="explorer-label-intro">attack:</span> <span
id="noscam-attack-label">-</span></span>
<span id="noscam-attack-score" class="score-value">-</span>
</div>
<div class="progress-bar">
<div id="noscam-attack-bar" class="bar-fill"></div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
<div class="navigation-controls">
<button id="prev-example" class="button nav-button">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"
style="vertical-align: text-bottom; margin-right: 5px;">
<line x1="19" y1="12" x2="5" y2="12"></line>
<polyline points="12 19 5 12 12 5"></polyline>
</svg>
Prev.
</button>
<div class="example-counter">
<span id="current-example-num">-</span> / <span id="total-examples">-</span>
</div>
<button id="next-example" class="button nav-button">
Next
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"
style="vertical-align: text-bottom; margin-left: 5px;">
<line x1="5" y1="12" x2="19" y2="12"></line>
<polyline points="12 5 19 12 12 19"></polyline>
</svg>
</button>
</div>
</div>
</section>
<section class="abstract">
<div class="container">
<h2>Abstract</h2>
<p>Typographic attacks exploit the interplay between text and visual content in multimodal foundation
models,
causing misclassifications when misleading text is embedded within images. Existing datasets are limited
in size and diversity, making it difficult to study such vulnerabilities. In this paper, we introduce
SCAM,
the largest and most diverse dataset of real-world typographic attack images to date, containing 1162
images
across hundreds of object categories and attack words.</p>
<p>Through extensive benchmarking of Vision-Language Models on SCAM, we demonstrate that typographic attacks
significantly degrade performance, and identify that training data and model architecture influence the
susceptibility to these attacks. Our findings indicate that typographic attacks remain effective against
state-of-the-art Large Vision-Language Models, especially those employing vision encoders inherently
vulnerable
to such attacks. However, employing larger Large Language Model backbones reduces this vulnerability
while
simultaneously enhancing typographic understanding. Additionally, we demonstrate that synthetic attacks
closely
resemble real-world (handwritten) attacks, validating their use in research. Our work provides a
comprehensive
resource and empirical insights to facilitate future research toward robust and trustworthy multimodal
AI systems.</p>
<h3>Key Contributions</h3>
<ul>
<li>Introduction of SCAM, the largest real-world typographic attack dataset to date</li>
<li>Comprehensive evaluation of VLMs and LVLMs, revealing an average drop in accuracy of 26%</li>
<li>Evidence that synthetic attacks closely resemble real-world attacks, validating their use in
research</li>
<li>Finding that LVLMs inherit vulnerabilities from vision encoders, but larger LLM backbones mitigate
weaknesses</li>
</ul>
</div>
</section>
<section class="dataset">
<div class="container">
<h2>SCAM Dataset</h2>
<p>SCAM is the largest and most diverse real-world typographic attack dataset to date, containing
images across hundreds of object categories and attack words.</p>
<div class="dataset-stats">
<div class="stat-box">
<div class="stat">
<h4>1,162</h4>
<p>Data points</p>
</div>
<div class="stat">
<h4>660</h4>
<p>Distinct object labels</p>
</div>
<div class="stat">
<h4>206</h4>
<p>Unique attack words</p>
</div>
<!-- <div class="stat">
<h4>1,147</h4>
<p>Unique object-word combinations</p>
</div> -->
</div>
<div class="dataset-comparison">
<h4>Attack Word Categories</h4>
<div class="comparison-image">
<img src="images/category_proportions.svg"
alt="Distribution of attack words in SCAM into categories">
</div>
<p>Distribution of attack words in SCAM into categories, highlighting both everyday terms and
safety-critical vocabulary.</p>
</div>
</div>
</div>
</section>
<section class="methodology">
<div class="container">
<h2>Evaluation Methodology</h2>
<div class="methodology-container">
<div class="methodology-vlm">
<h3>VLM Evaluation</h3>
<img src="images/method_VLM.svg" alt="VLM Evaluation Methodology">
<p>We evaluate the performance of VLMs in a zero-shot classification task. For each image, we
compute the cosine similarity between its embedding and the text embeddings of both the object
label and the attack word. The predicted label is determined based on the highest cosine
similarity score.</p>
</div>
<div class="methodology-lvlm">
<h3>LVLM Evaluation</h3>
<img src="images/method_LVLM.svg" alt="LVLM Evaluation Methodology">
<p>To assess the robustness of an LVLM against typographic attacks, we evaluate whether its output
changes when exposed to typographic modifications. Our evaluation involves providing the model
with an image along with a simple prompt asking it to identify what entity is depicted in the
image with two options.</p>
</div>
</div>
</div>
</section>
<section class="results">
<div class="container">
<h2>Results</h2>
<div class="result-section">
<h3>Impact on Vision-Language Models</h3>
<div class="result-figure">
<img src="images/all_evals_together.svg" alt="VLM Results" style="max-height: 250px;">
</div>
<p>Accuracy distribution of 99 VLMs across NoSCAM, SCAM, and SynthSCAM datasets. VLMs experience an
average accuracy drop of 26% when evaluated on the SCAM dataset, with an even steeper decline of 35%
on SynthSCAM.</p>
</div>
<div class="result-section">
<h3>Impact on Large Vision-Language Models and Dependence on Prompt</h3>
<div class="result-figure">
<img src="images/lvlm_prompt_evals.svg" alt="LVLM Results" style="max-height: 550px;">
</div>
<p>Smaller LLaVA models suffer substantial accuracy drops of 30-50%, while models with larger LLM
backbones exhibit better performance under attack.
Further, we evaluate whether prompting LVLMs to ignore the attack and focus on the object improves
robustness. While we cannot rule out the existence of an effective prompt, our results suggest it is
not effective.</p>
</div>
<div class="result-section">
<h3>Impact of Model Parameters</h3>
<div class="result-figure">
<img src="images/model_mparams_vs_lost_to_attack.svg" alt="Model Parameters vs Performance Drop"
style="max-height: 250px;">
</div>
<p>Susceptibility to typographic attack is agnostic of VLMs size, measured in millions of parameters.
While model size alone does not correlate with robustness, larger LLM backbones in LVLMs help
mitigate vulnerability to typographic attacks.</p>
</div>
<div class="result-section">
<h3>Impact of Attack Size</h3>
<div class="result-figure">
<img src="images/postit_size_vs_object_win_ratio.svg" alt="Post-it Size vs Object Win Ratio"
style="max-height: 300px;">
</div>
<p>Model accuracy decreases as the post-it area increases. The size of the attack text correlates with
model accuracy, with larger post-it notes causing greater performance degradation.</p>
</div>
</div>
</section>
<section class="key-findings">
<div class="container">
<h2>Key Findings</h2>
<div class="findings">
<div class="finding">
<div class="icon">📈</div>
<h3>Performance Impact</h3>
<p>Typographic attacks cause an average accuracy drop of 26% in VLMs and up to 50% in smaller
LLaVA models.</p>
</div>
<div class="finding">
<div class="icon">👁️</div>
<h3>Vision Encoder Vulnerability</h3>
<p>LVLMs inherit vulnerability to typographic attacks from their vision encoders, particularly the
ViT-L-14-336 backbone.</p>
</div>
<div class="finding">
<div class="icon">🧠</div>
<h3>LLM Backbone Effect</h3>
<p>Larger LLM backbones can compensate for vision encoder limitations, making models more resilient
to typographic attacks.</p>
</div>
<div class="finding">
<div class="icon">🔄</div>
<h3>Synthetic Validity</h3>
<p>Synthetic typographic attacks closely mirror real-world scenarios, validating their use for
evaluating model robustness.</p>
</div>
</div>
</div>
</section>
<section class="resources">
<div class="container">
<h2>Resources</h2>
<div class="resource-links">
<div class="resource">
<div class="icon">
<svg xmlns="http://www.w3.org/2000/svg" width="48" height="48" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
<path d="M14 2H6a2 2 0 0 0-2 2v16a2 2 0 0 0 2 2h12a2 2 0 0 0 2-2V8z"></path>
<polyline points="14 2 14 8 20 8"></polyline>
<line x1="16" y1="13" x2="8" y2="13"></line>
<line x1="16" y1="17" x2="8" y2="17"></line>
<polyline points="10 9 9 9 8 9"></polyline>
</svg>
</div>
<h3>Paper</h3>
<a href="https://arxiv.org/abs/2504.04893" class="button">arXiv</a>
</div>
<div class="resource">
<div class="icon">
<svg xmlns="http://www.w3.org/2000/svg" width="48" height="48" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
<path
d="M9 19c-5 1.5-5-2.5-7-3m14 6v-3.87a3.37 3.37 0 0 0-.94-2.61c3.14-.35 6.44-1.54 6.44-7A5.44 5.44 0 0 0 20 4.77 5.07 5.07 0 0 0 19.91 1S18.73.65 16 2.48a13.38 13.38 0 0 0-7 0C6.27.65 5.09 1 5.09 1A5.07 5.07 0 0 0 5 4.77a5.44 5.44 0 0 0-1.5 3.78c0 5.42 3.3 6.61 6.44 7A3.37 3.37 0 0 0 9 18.13V22">
</path>
</svg>
</div>
<h3>Code</h3>
<a href="https://github.com/Bliss-e-V/SCAM" class="button">GitHub Repository</a>
</div>
<div class="resource">
<div class="icon">
<svg xmlns="http://www.w3.org/2000/svg" width="48" height="48" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
<ellipse cx="12" cy="5" rx="9" ry="3"></ellipse>
<path d="M21 12c0 1.66-4 3-9 3s-9-1.34-9-3"></path>
<path d="M3 5v14c0 1.66 4 3 9 3s9-1.34 9-3V5"></path>
</svg>
</div>
<h3>Dataset</h3>
<a href="https://huggingface.co/datasets/BLISS-e-V/SCAM" class="button">Hugging Face</a>
</div>
<div class="resource">
<div class="icon">
<img src="images/poster.jpg" alt="SCAM Poster" style="height: 64px; margin: 0 auto;">
</div>
<h3>Poster</h3>
<a href="SCAM_CVPR_FoMo_poster.pdf" class="button">Download</a>
</div>
</div>
</div>
</section>
<section class="citation">
<div class="container">
<h2>Citation</h2>
<div class="citation-box">
<pre id="citation-text">@article{westerhoff2026scam,
title={{SCAM}: A Real-World Typographic Robustness Evaluation for Multimodal Foundation Models},
author={Justus Westerhoff and Erblina Purelku and Jakob Hackstein and Jonas Loos and Leo Pinetzki and Erik Rodner and Lorenz Hufe},
journal={Journal of Data-centric Machine Learning Research},
year={2026},
url={https://openreview.net/forum?id=zCcJSErVHH},
note={Dataset Certification}
}</pre>
<button id="copy-citation" class="button">
<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" viewBox="0 0 24 24" fill="none"
stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"
style="vertical-align: text-bottom; margin-right: 5px;">
<rect x="9" y="9" width="13" height="13" rx="2" ry="2"></rect>
<path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"></path>
</svg>
Copy
</button>
</div>
</div>
</section>
<footer>
<div class="container">
<p>© 2025 BLISS e.V.</p>
</div>
</footer>
<script src="script.js"></script>
</body>
</html>