generated from eliahuhorwitz/Academic-project-page-template
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
340 lines (303 loc) · 16.1 KB
/
index.html
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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<!-- Meta tags for social media banners, these should be filled in appropriatly as they are your "business card" -->
<!-- Replace the content tag with appropriate information -->
<meta name="description" content="DESCRIPTION META TAG">
<meta property="og:title" content="SOCIAL MEDIA TITLE TAG"/>
<meta property="og:description" content="SOCIAL MEDIA DESCRIPTION TAG TAG"/>
<meta property="og:url" content="URL OF THE WEBSITE"/>
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X630-->
<meta property="og:image" content="static/image/your_banner_image.png" />
<meta property="og:image:width" content="1200"/>
<meta property="og:image:height" content="630"/>
<meta name="twitter:title" content="TWITTER BANNER TITLE META TAG">
<meta name="twitter:description" content="TWITTER BANNER DESCRIPTION META TAG">
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X600-->
<meta name="twitter:image" content="static/images/your_twitter_banner_image.png">
<meta name="twitter:card" content="summary_large_image">
<!-- Keywords for your paper to be indexed by-->
<meta name="keywords" content="KEYWORDS SHOULD BE PLACED HERE">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Safety Alignment in NLP Tasks: Weakly Aligned Summarization as an In-Context Attack</title>
<link rel="icon" type="image/x-icon" href="static/images/cute-llama2.png">
<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 src="https://documentcloud.adobe.com/view-sdk/main.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>
</head>
<body>
<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">Safety Alignment in NLP Tasks: <br>Weakly Aligned Summarization as an In-Context Attack</h1>
<div class="is-size-5 publication-authors">
<!-- Paper authors -->
<span class="author-block">
<a href="https://fyyfu.github.io/" target="_blank">Yu Fu</a><sup>*</sup>,</span>
<span class="author-block">
<a href="https://yul091.github.io/liyufei.me/" target="_blank">Yufei Li</a><sup>*</sup>,</span>
<span class="author-block">
<a href="https://wendy-xiao.github.io/" target="_blank">Wen Xiao</a><sup>#</sup>,</span>
<span class="author-block">
<a href="https://intra.ece.ucr.edu/~cong/" target="_blank">Cong Liu</a><sup>*</sup>,</span>
<span class="author-block">
<a href="https://yuedong.us/" target="_blank">Yue Dong</a><sup>*</sup></span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block">University of California, Riverside<sup>*</sup>, Microsoft<sup>#</sup><br></span>
<!--span class="eql-cntrb"><small><br><sup>*</sup>Indicates Equal Contribution</small></span-->
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- Arxiv PDF link -->
<span class="link-block">
<a href="https://arxiv.org/pdf/2312.06924.pdf" target="_blank"
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>
<!-- Supplementary PDF link -->
<!-- span class="link-block">
<a href="static/pdfs/supplementary_material.pdf" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Supplementary</span>
</a>
</span-->
<!-- Github link -->
<span class="link-block">
<a href="https://github.com/FYYFU/SafetyAlignNLP/tree/master/data" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Data</span>
</a>
</span>
<!-- ArXiv abstract Link -->
<span class="link-block">
<a href="https://arxiv.org/abs/2312.06924" target="_blank"
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>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Teaser video-->
<!--<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<video poster="" id="tree" autoplay controls muted loop height="100%"> -->
<!-- Your video here -->
<!--
<source src="static/videos/banner_video.mp4"
type="video/mp4">
</video>
<h2 class="subtitle has-text-centered">
Aliquam vitae elit ullamcorper tellus egestas pellentesque. Ut lacus tellus, maximus vel lectus at, placerat pretium mi. Maecenas dignissim tincidunt vestibulum. Sed consequat hendrerit nisl ut maximus.
</h2>
</div>
</div>
</section> -->
<!-- End teaser video -->
<!-- Paper abstract -->
<!-- <section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
Recent developments in balancing the usefulness and safety of Large Language Models (LLMs) have raised a critical question: Are mainstream NLP tasks adequately aligned with safety considerations? Our study, focusing on safety-sensitive documents obtained through adversarial attacks, reveals significant disparities in the safety alignment of various NLP tasks. For instance, LLMs can effectively summarize malicious long documents but often refuse to translate them. This discrepancy highlights a previously unidentified vulnerability: attacks exploiting tasks with weaker safety alignment, like summarization, can potentially compromise the integrity of tasks traditionally deemed more robust, such as translation and question-answering (QA). Moreover, the concurrent use of multiple NLP tasks with lesser safety alignment increases the risk of LLMs inadvertently processing harmful content. We demonstrate these vulnerabilities in various safety-aligned LLMs, particularly Llama2 models and GPT-4, indicating an urgent need for strengthening safety alignments across a broad spectrum of NLP tasks.
</p>
</div>
</div>
</div>
</div>
</section> -->
<section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<!-- Column for Image -->
<div class="column is-two-fifths">
<img src="static/images/case.png" alt="Description of Image" />
</div>
<!-- Column for Text -->
<div class="column is-four-fifths">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
Recent developments in balancing the usefulness and safety of Large Language Models (LLMs) have raised a critical question: Are mainstream NLP tasks adequately aligned with safety considerations? Our study, focusing on safety-sensitive documents obtained through adversarial attacks, reveals significant disparities in the safety alignment of various NLP tasks. For instance, LLMs can effectively summarize malicious long documents but often refuse to translate them. This discrepancy highlights a previously unidentified vulnerability: attacks exploiting tasks with weaker safety alignment, like summarization, can potentially compromise the integrity of tasks traditionally deemed more robust, such as translation and question-answering (QA). Moreover, the concurrent use of multiple NLP tasks with lesser safety alignment increases the risk of LLMs inadvertently processing harmful content. We demonstrate these vulnerabilities in various safety-aligned LLMs, particularly Llama2 models and GPT-4, indicating an urgent need for strengthening safety alignments across a broad spectrum of NLP tasks. </p>
</div>
</div>
</div>
</div>
</section>
<!-- End paper abstract -->
<section class="section hero is-small">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full">
<div class="content">
<h2 class="title is-3">Method Overview</h2>
<center>
<img src="static/images/main.png" alt="Safety Alignment NLP" class="center-image blend-img-background"/>
</center>
<div class="level-set has-text-justified">
<p>
Our Contributions are:
<ol>
<li><strong>NLP Tasks Have Different Levels of Safety Alignment:</strong> We designed a novel setup using NLP task prompts and safety-sensitive documents, creating a dataset of 6,985 articles (average length of 1520 tokens) from adversarial attacks, to test whether different NLP tasks have varying levels of safety alignment. We found that tasks like summarization have notably lower safety alignment compared to translation or QA tasks. <br>
<li><strong>Weakly Aligned NLP Tasks as In-Context Attacks:</strong> The varying safety alignments among NLP tasks present a vulnerability. We discovered that performing weakly aligned NLP task first increases the likelihood of LLMs processing safety-sensitive documents for other tasks. This effect is further amplified when combining multiple weakly-aligned tasks. <br>
<li><strong>Vulnerability Cause Investigation:</strong> Our experiments indicate that safety alignment discrepancies in NLP tasks stem from an imbalanced trade-off between the usefulness from instruction tuning and the safety of alignment. Our ablation study reveals that summarization attacks are more frequently blocked on shorter documents than longer ones, possibly due to a prevalence of shorter documents in safety alignment. These findings are crucial for enhancing safety alignment research and building stronger defenses.
</ol>
</p>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section hero is-small">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full">
<div class="content">
<h2 class="title is-3">Task Process Rate Result on Safety-Sentitive Documents</h2>
<center>
<img src="static/images/radio-merge-use.png" alt="Safety Alignment NLP" class="center-image blend-img-background"/>
</center>
<div class="level-set has-text-justified">
<p>
Main results: NLP tasks have different levels of safety alignment on safety-sensitive (SS) documents. These safety-sensitive documents are obtained by adversarial attacks on LLMs based on malicious prompt from AdvBench.
</p>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="hero is-small is-light">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title is-3"> In-Context Attack with Summarization on GPT4</h2>
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<video id="replay-video"
controls
muted
preload
playsinline
height="100%">
<source src="./static/videos/GPT4-example.mp4"
type="video/mp4">
</video>
</center>
<div class="level-set has-text-justified">
<p>
In this example, GPT-4 initially refuses to translate the SS document concerning abuse. However, after processing a summarization first, it changes its mind and becomes willing to translate the document. In this case, the summarization task serves as an in-context attack, weakening the safety alignment of the translation task.
</p>
</div>
</div>
</div>
</div>
</div>
</section> -->
<!-- End youtube video
<!-- Video carousel -->
<section class="hero is-small">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title is-3">In-Context Attack with Summarization on Llama2</h2>
<div id="results-carousel" class="carousel results-carousel">
<div class="item item-video1">
<video poster="" id="video1" autoplay controls muted height="70%">
<!-- Your video file here -->
<source src="static/videos/llama2-13b-example1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-video2">
<video poster="" id="video2" autoplay controls muted height="70%">
<!-- Your video file here -->
<source src="static/videos/llama2-13b-example2.mp4"
type="video/mp4">
</video>
</div>
</div>
</div>
</div>
</section>
<!-- End video carousel -->
<!-- Paper poster -->
<!-- <section class="hero is-small is-light">
<div class="hero-body">
<div class="container">
<h2 class="title">Poster</h2>
<iframe src="static/pdfs/main.pdf" width="100%" height="550">
</iframe>
</div>
</div>
</section> -->
<!--End paper poster -->
<!--BibTex citation -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@misc{fu2023safety,
title={Safety Alignment in NLP Tasks: Weakly Aligned Summarization as an In-Context Attack},
author={Yu Fu and Yufei Li and Wen Xiao and Cong Liu and Yue Dong},
year={2023},
eprint={2312.06924},
archivePrefix={arXiv},
primaryClass={cs.CL}
}</code></pre>
</div>
</section>
<!--End BibTex citation -->
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This page was built using the <a href="https://github.com/eliahuhorwitz/Academic-project-page-template" target="_blank">Academic Project Page Template</a> which was adopted from the <a href="https://nerfies.github.io" target="_blank">Nerfies</a> project page.
</p>
</div>
</div>
</div>
</div>
</footer>
<!-- Statcounter tracking code -->
<!-- You can add a tracker to track page visits by creating an account at statcounter.com -->
<!-- End of Statcounter Code -->
</body>
</html>