BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//7.4.0.1//EN
TZID:Asia/Kolkata
X-WR-TIMEZONE:Asia/Kolkata
BEGIN:VEVENT
UID:213@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20260715T163000
DTEND;TZID=Asia/Kolkata:20260715T173000
DTSTAMP:20260709T161318Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-defense-cds-the-a
 rt-of-control-post-hoc-alignment-of-diffusion-models-for-safety-ethics-and
 -fine-grained-control/
SUMMARY:M.Tech Research Thesis Defense}: CDS: "The Art of Control: Post-hoc
  Alignment of Diffusion Models for Safety\, Ethics\, and Fine-grained Cont
 rol"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nM.Tech Research 
 Thesis Defense\n\n\n\nSpeaker : Mr. Aakash Kumar Singh\nS.R. Number : 06-1
 8-01-10-22-23-1-23884\nTitle : "The Art of Control: Post-hoc Alignment of 
 Diffusion Models for Safety\, Ethics\, and Fine-grained Control."\nThesis 
 examiner: Prof. Mayank Vatsa\, IIT Jodhpur\nResearch Supervisor : Prof. Ve
 nkatesh Babu\nDate &amp\; Time : July 15\, 2026\, 04.30 PM\nVenue : # 102 
 CDS Seminar Hall\n\n\n\nABSTRACT\n\nDiffusion models have achieved unprece
 dented success in generating high-fidelity images\, largely due to the mas
 sive scale of their training datasets. Yet this success raises significant
  concerns regarding privacy\, intellectual property rights\, and model saf
 ety as these methods tend to generate copyrighted content and accurate hum
 an identities seen during the training phase. To address these risks\, ear
 lier approaches relied on dataset filtering and inference-time safety chec
 kers—both fundamentally fragile solutions. These external defenses merel
 y suppress outputs without removing the underlying internal representation
 s\, leaving undesirable concepts intact.\n\nThese concerns highlight the n
 eed for unlearning concepts post-training. Existing unlearning methods can
  erase specific concepts but often introduce significant degradation in ne
 ighbouring concepts. To reduce such side effects\, recent approaches requi
 re extensive domain expertise to identify which other concepts must be pre
 served\, making them impractical for large-scale deployment where thousand
 s of concepts must be handled automatically. Moreover\, many sensitive con
 cepts (NSFW content) are inherently subjective\, with desired degree of fo
 rgetting varying across cultural\, regional\, and application-specific con
 texts\, further complicating the use of rigid\, one-size-fits-all unlearni
 ng strategies.\n\nTo address these challenges\, we introduce Concept Sieve
 r\, an end-to-end framework for targeted concept removal in pre-trained te
 xt-to-image diffusion models. Concept Siever rests on two key innovations:
 \n\n1. Automatic Paired Data Generation : The framework creates paired dat
 asets of a target concept and its negations by utilizing the diffusion mod
 el's own latent space. These pairs differ primarily in the target concept\
 , enabling precise forgetting with reduced side effects—crucially\, with
 out requiring domain expertise.\n\n2. Concept-Specific Localization : Conc
 ept Siever employs a novel localization method to identify and isolate mod
 el components most responsible for the target concept. By retraining only 
 these localized components on the paired dataset\, the method accurately r
 emoves concepts with reduced side effects while preserving neighboring and
  unrelated concepts. Additionally\, Concept Siever offers continuous infer
 ence-time control over forgetting strength\, enabling flexible adaptation 
 to context-dependent requirements without additional fine-tuning. The meth
 od achieves state-of-the-art performance on the I2P benchmark\, surpassing
  previous methods by over 33\\% while demonstrating superior structure pre
 servation. Extensive quantitative and qualitative evaluations\, along with
  a user study\, validate the effectiveness of our approach in providing a 
 targeted\, adjustable mechanism for concept erasure with reduced collatera
 l impact\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VTIMEZONE
TZID:Asia/Kolkata
X-LIC-LOCATION:Asia/Kolkata
BEGIN:STANDARD
DTSTART:20250715T163000
TZOFFSETFROM:+0530
TZOFFSETTO:+0530
TZNAME:IST
END:STANDARD
END:VTIMEZONE
END:VCALENDAR