Yihua Zhang
@zyh2022
Followers
422
Following
371
Media
28
Statuses
145
Ph.D. student in CS at @michiganstateu. Trustworthy and Scalable AI. IBM PhD Fellowship Awardee 2024, Previously @Amazon, @Cisco, and currently @MetaAI.
East Lansing
Joined February 2022
RT @zhoubolei: @NeurIPSConf This is great! But will you also consider setting up an official satellite location in China, given the fact th….
0
6
0
RT @ChongyuFan: Excited to share our ICML '25 work on robust LLM unlearning—“Towards LLM Unlearning Resilient to Relearning Attacks: A Shar….
0
2
0
Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.
368
648
3K
RT @sijialiu17: 🚨 Excited to attend #ICML2025 and share our latest work (@OptML_MSU) on LLM unlearning -- think of it as AI surgery: removi….
0
12
0
RT @pinyuchenTW: (2/n) Oral Paper!.Task vector is one of the most popular methods for model editing and model merging. We provide a theoret….
0
1
0
RT @sijialiu17: Check out our oral paper #ICLR25, led by @LiHongkang_jntm and @zyh2022, which shows when and why model editing via task vec….
0
4
0
I'm thrilled to announce that I've been awarded the prestigious IBM PhD Fellowship 2024! @IBMResearch. A heartfelt thank you to my advisor, colleagues, and the @IBM award committee for their support and recognition. #IBMPhDFellowship2024.
research.ibm.com
IBM Announces 2024 PhD Fellowship Awardees to research AI, Cloud, Data Science, Open Source, Quantum and Security.
4
0
22
So grateful for such a babysitting-style tutorial! It took me ten minutes to read but I guess it will take me several months to digest. Thanks Manling!.
If you are asked "Do you have any questions for me?" during the interview, here are some I like to ask:. Research:.- How do faculty members collaborate? Do you often apply for funding together?.- Is co-advising students a common practice here? .- Are there any plans or proposals.
0
0
2
🎉🎉🎉Who needs Santa when your boss is delivering wins like this? 🌟🌟🌟.
Honored to receive the Amazon Research Award for advancing machine unlearning techniques--what a fantastic Christmas gift from @AmazonScience! Huge thanks to my hardworking students @OptML_MSU. Looking forward to new and fruitful collaborations with Amazon in 2025! @MSU_EGR.
0
0
1
RT @sijialiu17: Honored to receive the Amazon Research Award for advancing machine unlearning techniques--what a fantastic Christmas gift f….
0
2
0
🚨 Call to Action: Demand Accountability and Request Refunds from NeurIPS! 🚨. 🛑 To @NeurIPSConf: If you don't show your attitude, we will show ours! 🛑. We demand: .1️⃣ A full public apology from NeurIPS and the keynote speaker. 2️⃣ Refunds for all affected attendees. ❗️ Let’s
7
14
210
“I thought adding some vague and evasive NOTE would make my racially discriminatory behavior seem less offensive.”. - Excuse given by MIT professor who is scheming and racist and now in the eye of the storm. NOTE: Most MIT professors who I know are not discriminative or racist.
Mitigating racial bias from LLMs is a lot easier than removing it from humans! . Can’t believe this happened at the best AI conference @NeurIPSConf . We have ethical reviews for authors, but missed it for invited speakers? 😡
0
2
27
RT @sijialiu17: Excited to attend #NeurIPS2024 in Vancouver! 🎉. Check out the OPTML Menu of Innovations @ NeurIPS 2024:.🔗 .
0
6
0
🙌 Acknowledgments. Thanks to my collaborators: @hhwang1108 , @pkbye97, Prof. @y1923z01a, Prof. @sijialiu17 and Prof. @_vztu for their contributions! Special thanks to Hanhui, my summer intern this year, who is currently applying for a Ph.D. program starting in Fall 2025. He is.
0
0
1
🎉 Bonus: Vulnerabilities in current editing evaluation metrics.The evaluation of photo and video editing tasks also has vulnerabilities:.⚠️ CLIP-based scores can encourage ‘over-editing,’ prioritizing extreme changes over content quality. ⚠️ SSIM/PSNR metrics often rely heavily
1
0
0
📝 Postscript. FACELOCK is not without its limitations, but it opens up a new technical direction, raising awareness of alternative approaches to photo protection. We envision a future where social platforms offer users the option to protect their photos before posting, defending.
1
0
0
📈 Results. Compared to existing methods like PhotoGuard and EditShield, FACELOCK demonstrates superior protection and robustness across various scenarios:.✔️ It performs significantly better against Gaussian noise, rotation, JPEG compression, and other perturbations. ✔️ It
1
0
1
🛡️ Adversarial perturbations to make biometrics fragile. It turns out that erasing biometric information is no easy task. Diffusion models naturally purify small adversarial perturbations during their editing process. To overcome this, we integrated the following three levels
1
0
0
🔍 Insights from the perspective of malicious editors. Let’s consider the problem from the perspective of a malicious editor: What constitutes a ‘successful edit’?.Take a photo of Trump, for example. If the goal is to make him ‘wear a police uniform,’ a successful edit must
1
1
1
🤔 Is making photos ‘uneditable’ the only viable protection path?. Using adversarial perturbations—imperceptible pixel-level noise generated through optimization—to make photos ‘uneditable’ has been a mainstream approach. However, this method has significant limitations:.📌 The
1
0
0