1<!DOCTYPE HTML> 2<html lang="en"> 3 <head> 4 5 6<!-- Google tag (gtag.js) -->
vendor: 65 bytes, lines 6-7
6 7 <script async src="https://www.googletagmanager.com/gtag/js?id=
7G-EVNJFH5QFJ
vendor: 13 bytes, lines 7-8
7"></script> 8
8<script> 9
vendor: 138 bytes, lines 9-13
9window.dataLayer = window.dataLayer || []; 10 function gtag(){dataLayer.push(arguments);} 11 gtag('js', new Date()); 12 13 gtag('config', '
13G-EVNJFH5QFJ
vendor: 5 bytes, lines 13-14
13'); 14
14</script>
14 15 16 17 18 <meta http-equiv="Content-Type" content="text/html; charset=UTF-8"> 19 20 <title>Muhammad Huzaifa</title> 21 22 <meta name="author" content="Muhammad Huzaifa"> 23 <meta name="viewport" content="width=device-width, initial-scale=0.4"> 24 <link rel="shortcut icon" href="data/favicon/favicon.ico" type="image/x-icon"> 25 <link rel="stylesheet" type="text/css" href="stylesheet.css"> 26 27 </head> 28 29 30 31 <body> 32 <table style="width:100%;max-width:800px;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 33 <tr style="padding:0px"> 34 <td style="padding:0px"> 35 <table style="width:100%;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 36 <tr style="padding:0px"> 37 <td style="padding:2.5%;width:63%;vertical-align:middle"> 38 <p class="name" style="text-align: center;"> 39 Muhammad Huzaifa 40 </p> 41 <p> 42 I am a doctoral researcher at the 43 <a href="https://cispa.de/en" target="_blank" rel="noopener noreferrer"> 44 CISPA Helmholtz Center for Information Security 45 </a>, 46 where I work closely with 47 <a href="https://profiles.cispa.de/c01thei" target="_blank" rel="noopener noreferrer"> 48 Thorsten Eisenhofer 49 </a> 50 and 51 <a href="https://profiles.cispa.de/c01lesc" target="_blank" rel="noopener noreferrer"> 52 Lea Schönherr 53 </a>. 54 I am passionate about adversarial machine learning, with the goal of developing AI systems that extend human intelligence in a secure and privacy-preserving manner. 55 My current research focuses on the safety and security of generative AI. 56 </p> 57 <p> 58 Prior to my current role, I completed my Masterâs degree at 59 <a href="https://mbzuai.ac.ae/" target="_blank" rel="noopener noreferrer"> 60 Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) 61 </a> 62 under the supervision of 63 <a href="https://mbzuai.ac.ae/study/faculty/fahad-khan/" target="_blank" rel="noopener noreferrer"> 64 Prof. Fahad Khan 65 </a> 66 and 67 <a href="https://mbzuai.ac.ae/study/faculty/salman-khan/" target="_blank" rel="noopener noreferrer"> 68 Prof. Salman Khan 69 </a>. 70 During this time, my research spanned multimodal representation learning, the robustness of vision models, and the adaptation of vision-language models to low-data settings, including few-shot and zero-shot scenarios. 71 Before that, I earned my undergraduate degree in Electrical Engineering from NUST, Islamabad. 72 Along the way, I also worked as a researcher at the 73 <a href="https://tuebingen.ai/" target="_blank" rel="noopener noreferrer"> 74 Tübingen AI Center 75 </a> 76 and 77 <a href="https://mbzuai.ac.ae/" target="_blank" rel="noopener noreferrer"> 78 MBZUAI 79 </a>. 80 <p style="margin-top: 12px; margin-bottom: 0px; color: #287489;"> 81 I am always open to collaborations and discussions on research, security, privacy, or related topics. Feel free to 82 <a href="mailto:[email protected]" style="color: #2d23ea;">email me</a>. 83 </p> 84 <p style="text-align:center"> 85 <a href="mailto:[email protected]">Email</a> / 86 <a href="https://www.linkedin.com/in/muhammad-huzaifa--/">Linkedin</a> / 87 <a href="https://github.com/Muhammad-Huzaifaa">Github</a> / 88 <a href="https://scholar.google.com/citations?user=V7hTDxQAAAAJ&hl=en">Scholar</a> / 89 <!-- <a href="data/Resume.pdf">CV</a> / --> 90 <a href="https://twitter.com/huzaifa2k50">Twitter</a> / 91 <a href="https://bsky.app/profile/mhuxaifa.bsky.social">Bluesky</a> 92 </p> 93 </td> 94 <td style="padding:2.5%;width:26%;max-width:26%"> 95 <a href="data/profile_pic.jpg"> 96 <img 97 src="data/profile_pic.jpg" 98 alt="profile photo" 99 style=" 100 width: 100%; 101 height: auto; 102 max-width: 320px; 103 aspect-ratio: 1 / 1; 104 object-fit: cover; 105 object-position: center 20%; /* anchor crop higher on the face */ 106 transform: scale(1.25); /* zoom in 25% */ 107 transform-origin: center 30%; /* zoom toward face area */ 108 border-radius: 50%; 109 box-shadow: 0 10px 30px rgba(0, 0, 0, 0.18); 110 display: block; 111 margin-left: auto; 112 overflow: hidden; 113 " 114 class="hoverZoomLink" 115 > 116 </a> 117 </td> 118 </tr> 119 </tbody></table> 120 121 <table align="center" border="0" cellspacing="0" cellpadding="0" style="width:100%;max-width:1000px;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto; margin-top: -10px"><tbody> 122
123 <!-- <table width="100%" align="center" border="0" cellspacing="0" cellpadding="20" > --> 124 <tr> 125 <td width="20%" valign="middle"> 126 <a href="https://cispa.de/en"><img src="data/cispa.png" width="100"></a> 127 </td> 128 <td width="20%" valign="middle"> 129 <a href="https://www.uni-saarland.de/en/home.html"><img src="data/uds.png" width="100"></a> 130 </td> 131 <td width="20%" valign="middle"> 132 <a href="https://tuebingen.ai/"><img src="data/ai_center.png" width="100"></a></td> 133 </td> 134 <td width="20%" valign="middle"> 135 <a href="https://mbzuai.ac.ae/"><img src="data/mbzuai.png" width="100"></a></td> 136 </td> 137 <td width="20%" valign="middle"> 138 <a href="https://nust.edu.pk/"><img src="data/NUST-logo.png" width="100"></a></td> 139 </td> 140 </tr> 141 </table> 142 143 144 <table style="width:100%;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 145 146 147 148 <tr> 149 <td colspan="2" style="padding:16px 2.5% 0 2.5%;"> 150 <h2 style="margin:0 0 10px 0;">News</h2> 151 152 <div class="news-box"> 153 <div class="news-scroll"> 154 155 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2026-09]</span> Released <b>Safety of Latent Communication in Multi-Agent Systems</b> on <a href="https://arxiv.org/abs/2609.39788">arXiv</a>. Code coming soon.</div> 156 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2026-09]</span> Released <b>WISE-ATTA</b> on <a href="https://arxiv.org/abs/2609.37687">arXiv</a> with <a href="https://github.com/Muhammad-Huzaifaa/WISE-ATTA">code</a>.</div> 157 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2026-05]</span> Started my <b>PhD at CISPA</b>.</div> 158 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2026-04]</span> One paper accepted at the <b>GRAIL-V Workshop, CVPR 2026</b>.</div> 159 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2026-02]</span> One paper accepted to the <b>CVPR 2026 main track</b>.</div> 160 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2025-11]</span> Started a <b>Research Assistant</b> position at <b>CISPA</b> under <b>Prof. Thorsten Eisenhofer</b>.</div> 161 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2025-07]</span> Started a <b>Research Assistant</b> position at the <b>Tübingen AI Center</b> under <b>Prof. Hilde Kuehne</b>.</div> 162 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2025-07]</span> <b>TTA</b> accepted at <b>WACV 2025</b>.</div> 163 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2025-06]</span> <b>ObjectCompose</b> received an <b>Honorable Mention for Best Paper</b> at <b>ACCV 2024</b>.</div> 164 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2025-06]</span> <b>ObjectCompose</b> accepted at <b>ACCV 2024</b>.</div> 165 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2025-02]</span> <b>Defensive Diffusion</b> accepted at <b>MIUA 2023</b>.</div> 166 <div class="collapsible-header"><span style="color: #2563eb; font-weight: 600;">[2025-02]</span> Awarded a <b>fully funded Masterâs scholarship</b> at <b>MBZUAI</b>.</div> 167 168 </div> 169 </div> 170 </td> 171 </tr> 172 173 174 175 <!-- Selected Publications header --> 176 <tr> 177 <td colspan="2" style="padding:16px 2.5% 6px 2.5%;"> 178 <div style="display:flex; justify-content:space-between; align-items:baseline; gap:12px; flex-wrap:wrap;"> 179 <h2 style="margin:0;">Selected Publications</h2> 180 <!-- optional link --> 181 <a href="https://scholar.google.com/citations?user=V7hTDxQAAAAJ&hl=en" target="_blank" rel="noopener noreferrer">Full publication list â</a> 182 </div> 183 </td> 184 </tr> 185 186 <!-- Latent Communication Safety --> 187 <tr> 188 <td style="padding:16px;width:20%;vertical-align:middle"> 189 <div class="one"> 190 <img src="data/latent_safety.png" width="160" style="padding-top:30px;"> 191 </div> 192 </td> 193 <td style="padding:8px;width:80%;vertical-align:middle"> 194 <a href="https://arxiv.org/abs/2609.39788">
195 <span class="papertitle">Safety of Latent Communication in Multi-Agent Systems</span> 196 </a> 197 <br> 198 <strong>Muhammad Huzaifa</strong>, 199 Sina Mavali, 200 <a href="https://profiles.cispa.de/c01thei" target="_blank" rel="noopener noreferrer">Thorsten Eisenhofer</a> 201 <br> 202 <em>arXiv</em>, 2026 203 <br> 204 <a href="https://arxiv.org/abs/2609.39788" target="_blank" rel="noopener noreferrer">arXiv</a> / 205 <a href="https://github.com/Muhammad-Huzaifaa/latent-safety" target="_blank" rel="noopener noreferrer">code</a> 206 <p></p> 207 <p>We show that learned latent communication links can compromise the safety of multi-agent systems built from safety-aligned agents, and that updating the links alone can repair them.</p> 208 </td> 209 </tr> 210 211 <!-- WISE-ATTA --> 212 <tr> 213 <td style="padding:16px;width:20%;vertical-align:middle"> 214 <div class="one"> 215 <img src="data/wise_atta.png" width="160" style="padding-top:30px;"> 216 </div> 217 </td> 218 <td style="padding:8px;width:80%;vertical-align:middle"> 219 <a href="https://arxiv.org/abs/2609.37687"> 220 <span class="papertitle">WISE-ATTA: When to Ask for Labels in Budgeted Active Test-Time Adaptation</span> 221 </a> 222 <br> 223 <strong>Muhammad Huzaifa</strong>, 224 <a href="https://profiles.cispa.de/c01lesc" target="_blank" rel="noopener noreferrer">Lea Schönherr</a>, 225 <a href="https://profiles.cispa.de/c01thei" target="_blank" rel="noopener noreferrer">Thorsten Eisenhofer</a> 226 <br> 227 <em>arXiv</em>, 2026 228 <br> 229 <a href="https://arxiv.org/abs/2609.37687" target="_blank" rel="noopener noreferrer">arXiv</a> / 230 <a href="https://github.com/Muhammad-Huzaifaa/WISE-ATTA" target="_blank" rel="noopener noreferrer">code</a> 231 <p></p> 232 <p>WISE-ATTA decides online when to request labels under a fixed annotation budget, matching or outperforming recent active test-time adaptation methods with far fewer labels.</p> 233 </td> 234 </tr> 235 236 <!-- Paper 1 --> 237 238 <tr onmouseout="visualoverload_start()" onmouseover="visualoverload_start()"> 239 <td style="padding:16px;width:20%;vertical-align:middle"> 240 <div class="one"> 241 <img src="data/visual-overload.png" width="160" style="padding-top:30px;"> 242 </div> 243 </td> 244 <td style="padding:8px;width:80%;vertical-align:middle"> 245 <a href="https://paulgavrikov.github.io/visualoverload/"> 246 <span class="papertitle">VisualOverload: Probing Visual Understanding of VLMs in Really Dense Scenes</span> 247 </a> 248 <br> 249 <a href="https://scholar.google.com/citations?hl=en&user=cshJtrQAAAAJ" target="_blank" rel="noopener noreferrer">Paul Gavrikov</a>, 250 <a href="https://scholar.google.com/citations?hl=en&user=JJRr8c8AAAAJ" target="_blank" rel="noopener noreferrer">Wei Lin</a>, 251 <a href="https://scholar.google.com/citations?hl=en&user=cES2rkAAAAAJ" target="_blank" rel="noopener noreferrer">M. Jehanzeb Mirza</a>, 252 <a href="https://scholar.google.com/citations?hl=en&user=oA25i6QAAAAJ" target="_blank" rel="noopener noreferrer">Soumya Jahagirdar</a>, 253 <strong>Muhammad Huzaifa</strong>, 254 <a href="https://scholar.google.com/citations?hl=iw&user=ER4dt8cAAAAJ" target="_blank" rel="noopener noreferrer">Sivan Doveh</a>, 255 <a href="https://scholar.google.com/citations?hl=en&user=Tw2m5kUAAAAJ" target="_blank" rel="noopener noreferrer">Serena Yeung-Levy</a>, 256 <a href="https://scholar.google.com/citations?hl=en&user=pfGI-KcAAAAJ" target="_blank" rel="noopener noreferrer">James Glass</a>, 257 <a href="https://scholar.google.com/citations?hl=en&user=pxhCcH0AAAAJ" target="_blank" rel="noopener noreferrer">Hilde Kuehne</a> 258 <br> 259 <em>CVPR</em>, 2026 260 <br> 261 <a href="https://arxiv.org/abs/2509.25339" target="_blank" rel="noopener noreferrer">arXiv</a> / 262 <a href="https://github.com/paulgavrikov/visualoverload" target="_blank" rel="noopener noreferrer">code</a> / 263 <a href="https://paulgavrikov.github.io/visualoverload/" target="_blank" rel="noopener noreferrer">project page</a> / 264 <p></p> 265 <p>VisualOverload introduces a dense-scene benchmark for probing detailed visual understanding in vision-language models, revealing substantial performance gaps on crowded, high-detail images.</p> 266 </td> 267 </tr> 268 269 <!-- Paper 2 --> 270 271 <tr onmouseout="objectcompose_start()" onmouseover="objectcompose_start()"> 272 <td style="padding:16px;width:20%;vertical-align:middle"> 273 <div class="one"> 274 <img src="data/objectcompose.jpg" width="160" style="padding-top:30px;"> 275 </div> 276 </td> 277 <td style="padding:8px;width:80%;vertical-align:middle"> 278 <a href="https://muhammad-huzaifaa.github.io/ObjectC
278ompose/"> 279 <span class="papertitle">ObjectCompose: Evaluating Resilience of Vision-Based Models on Object-to-Background Compositional Changes</span> 280 </a> 281 <br> 282 <a href="https://scholar.google.com/citations?hl=en&user=2Ft7r4AAAAAJ" target="_blank" rel="noopener noreferrer">Hashmat Shadab Malik</a>*, 283 <strong>Muhammad Huzaifa</strong>*, 284 <a href="https://scholar.google.com/citations?hl=en&user=tM9xKA8AAAAJ" target="_blank" rel="noopener noreferrer">Muzammal Naseer</a>, 285 <a href="https://scholar.google.com/citations?hl=en&user=M59O9lkAAAAJ" target="_blank" rel="noopener noreferrer">Salman Khan</a>, 286 <a href="https://scholar.google.com/citations?hl=en&user=zvaeYnUAAAAJ" target="_blank" rel="noopener noreferrer">Fahad Shahbaz Khan</a> 287 <br> 288 <em >ACCV</em>, 2024 289 <strong style="color: #dc2626;">(Best Student Paper Honorable Mention)</strong> 290 <br> 291 <a href="https://arxiv.org/abs/2403.04701" target="_blank" rel="noopener noreferrer">arXiv</a> / 292 <a href="https://github.com/Muhammad-Huzaifaa/ObjectCompose" target="_blank" rel="noopener noreferrer">code</a> / 293 <a href="https://muhammad-huzaifaa.github.io/ObjectCompose/" target="_blank" rel="noopener noreferrer">project page</a> / 294 <p></p> 295 <p>ObjectCompose evaluates how robust vision-based models are to object-to-background compositional shifts, providing a benchmark for studying context sensitivity in modern visual recognition systems.</p> 296 </td> 297 </tr> 298 299 300<!-- 301 Paper 3 â EFSA (hidden) 302 <tr onmouseout="efsa_start()" onmouseover="efsa_start()"> 303 <td style="padding:16px;width:20%;vertical-align:middle"> 304 <div class="one"> 305 <img src="data/efsa.jpg" width="160" style="padding-top:30px;"> 306 </div> 307 </td> 308 <td style="padding:8px;width:80%;vertical-align:middle"> 309 <a href="https://arxiv.org/abs/2412.00139"> 310 <span class="papertitle">EFSA: Episodic Few-Shot Adaptation for Text-to-Image Retrieval</span> 311 </a> 312 <br> 313 <strong>Muhammad Huzaifa</strong>, 314 <a href="https://scholar.google.com/citations?hl=en&user=Br-FqfIAAAAJ" target="_blank" rel="noopener noreferrer">Yova Kementchedjhieva</a> 315 <br> 316 <em>CVPRW</em>, 2026 317 <br> 318 <a href="https://arxiv.org/abs/2412.00139" target="_blank" rel="noopener noreferrer">arXiv</a> / 319 <a href="https://github.com/Muhammad-Huzaifaa/EFSA" target="_blank" rel="noopener noreferrer">code</a> / 320 321 <p></p> 322 <p>EFSA proposes a test-time episodic few-shot adaptation framework for open-domain text-to-image retrieval, improving robustness across diverse query domains and hard negatives.</p> 323 </td> 324 </tr> 325--> 326 327 328 329 </tbody></table> 330 331 332 <table style="width:100%; margin:0 auto; border:0; border-spacing:0; padding:16px;"><tbody> 333 <tr> 334 <td> 335 <h2>Miscellanea</h2> 336 </td> 337 </tr> 338 </tbody></table> 339 <table style="width:100%;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 340 341 <!-- <tr> 342 <td align="center" style="padding:16px;width:20%;vertical-align:middle"> 343 <div class="colored-box" style="background-color: #aaba9e;"> 344 <h2>Experience</h2> 345 </div> 346 </td> 347 <td style="padding:8px;width:80%;vertical-align:center"> 348 <a href="https://tech.facebook.com/reality-labs/">Meta</a>, Research Intern - Meta Reality Labs. 349 <br> 350 <a href="https://www.logitech.com/en-ch">Logitech</a>, Research Intern - Immersive Video Conferencing. 351 <br> 352 <a href="https://www.agile-robots.com">Agile Robots AG</a>, Research Intern - Applied Machine Learning. 353 <br> 354 <a href="https://www.mpe.mpg.de/index">
354Max Planck Institute</a>, Research Intern - Deep Learning on FPGA. 355 <br> 356 <a href="https://www.bmwgroup.com/en/general/approach-autonomous-driving-campus.html">BMW Group, Autonomous Driving Campus</a>, Research Intern â HW/SW Optimization of CNNs. 357 <br> 358 <a href="https://www.gehealthcare.com">GE-Healthcare</a> Intern â Software Development and Testing. 359 <br> 360 <a href="https://www.siemens.com">Siemens AG</a> Research Intern â Deep Learning Model Deployment. 361 <br> 362 <a href="https://www.intel.com">Intel</a> Working Student â Software Development. 363 </td> 364 </tr> --> 365 366 <!-- <tr> 367 <td align="center" style="padding:16px;width:20%;vertical-align:middle"> 368 <div class="colored-box" style="background-color: #fcb97d;"> 369 <h2>Supervised Students</h2> 370 </div> 371 </td> 372 <td style="padding:8px;width:80%;vertical-align:middle"> 373 374 <a href="https://ckli.alphaxiv.io/">Chengkun Li</a>(Now PhD Student at EPFL) 375 [Modular Quantization for Object Detection.] 376 <br> 377 <a href="https://robertflame.github.io/Homepage/">Ziqi Zhao</a> (Now PhD Student at HKU). 378 [Modular Quantization for 6D pose.] 379 <br> 380 <a href="mailto:[email protected]">Ahmad Jarrar Khan</a>(Master Student at EPFL) 381 [Compressed Gaussians for Dynamic 3D scenes.] 382 <br> 383 <a href="https://www.aubep.ch/entreprise/equipe-aubep">Kepler Warrington-Arroyo</a> (Consultant, Aubep) 384 [Vision Language Model compression]. 385 <br> 386 </td> 387 </tr> --> 388 389 390 391 392 <tr> 393 <td align="center" style="padding:16px;width:20%;vertical-align:middle"> 394 <div class="colored-box" style="background-color: #c6b89e;"> 395 <h2>Academic Service</h2> 396 </div> 397 </td> 398 <td style="padding:8px;width:80%;vertical-align:center"> 399 <a href="https://cvpr.thecvf.com/">Reviewer, CVPR 2025, 2026</a> 400 <br> 401 <a href="https://eccv.ecva.net/">Reviewer, ECCV 2024, 2026</a> 402 <br> 403 <a href="https://neurips.cc/">Reviewer, NeurIPS 2026</a> 404 <br> 405 <a href="https://iccv2023.thecvf.com/">Reviewer, ICCV 2025</a> 406 <br> 407 <a href="https://wacv.thecvf.com/">Reviewer, WACV 2025, 2026</a> 408 <br> 409 </td> 410 </tr> 411 412 413 414 <tr> 415 <td align="center" style="padding:16px;width:20%;vertical-align:middle"> 416 <div class="colored-box" style="background-color: #edd892;"> 417 <h2>Teaching Assistant</h2> 418 </div> 419 </td> 420 <td style="padding:8px;width:80%;vertical-align:center"> 421 <a>Foundations of Artificial Intelligence AI701 (MBZUAI)</a>, Fall 2023. 422 <br> 423 <a>Introduction to Machine Learning ML701 (MBZUAI)</a>, Fall 2023. 424 <br> 425 <a>Deep Learning AI-702 (MBZUAI)</a>, Spring 2023. 426 </td> 427 </tr> 428 429 </tbody></table> 430 <table style="width:100%;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 431 <tr> 432 <!-- <td style="padding:0px"> 433 <br> 434 <p style="text-align:right;font-size:small;"> 435 <a href="https://github.com/jonbarron/jonbarron_website">Template source code</a>. 436 </p> 437 </td> --> 438 </tr> 439 </tbody></table> 440 </td> 441 </tr> 442 </table> 443 </body> 444</html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.