BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//7.2.3.1//EN
TZID:Asia/Kolkata
X-WR-TIMEZONE:Asia/Kolkata
BEGIN:VEVENT
UID:177@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20260116T103000
DTEND;TZID=Asia/Kolkata:20260116T113000
DTSTAMP:20260112T154039Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-colloquium-cds-th
 e-art-of-control-post-hoc-alignment-of-diffusion-models-for-safety-ethics-
 and-fine-grained-control/
SUMMARY:M.Tech Research Thesis {Colloquium}: CDS: "The Art of Control: Post
 -hoc Alignment of Diffusion Models for Safety\, Ethics\, and Fine-grained 
 Control"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nM.Tech Research 
 Thesis Colloquium\n\n\n\nSpeaker : Mr. Aakash Kumar Singh\nS.R. Number : 0
 6-18-01-10-22-23-1-23884\nTitle : " The Art of Control: Post-hoc Alignment
  of Diffusion Models for Safety\, Ethics\, and Fine-grained Control."\nRes
 earch Supervisor : Prof. Venkatesh Babu\nDate &amp\; Time : January 16\, 2
 026\, 10.30 AM\nVenue : # 102 CDS Seminar Hall\n\n\n\nABSTRACT\n\nDiffusio
 n models have achieved unprecedented success in generating high-fidelity i
 mages\, largely due to the massive scale of their training datasets. Yet t
 his success raises significant concerns regarding privacy\, intellectual p
 roperty rights\, and model safety as these methods tend to generate copyri
 ghted content and accurate human identities seen during the training phase
 . To address these risks\, earlier approaches relied on dataset filtering 
 and inference-time safety checkers—both fundamentally fragile solutions.
  These external defenses merely suppress outputs without removing the unde
 rlying internal representations\, leaving undesirable concepts intact.\n\n
 These concerns highlight the need for unlearning concepts post-training. E
 xisting unlearning methods can erase specific concepts but often introduce
  significant degradation in neighbouring concepts. To reduce such side eff
 ects\, recent approaches require extensive domain expertise to identify wh
 ich other concepts must be preserved\, making them impractical for large-s
 cale deployment where thousands of concepts must be handled automatically.
  Moreover\, many sensitive concepts (NSFW content) are inherently subjecti
 ve\, with desired degree of forgetting varying across cultural\, regional\
 , and application-specific contexts\, further complicating the use of rigi
 d\, one-size-fits-all unlearning strategies.\n\nTo address these challenge
 s\, we introduce Concept Siever\, an end-to-end framework for targeted con
 cept removal in pre-trained text-to-image diffusion models. Concept Siever
  rests on two key innovations:\n\n 	Automatic Paired Data Generation : The
  framework creates paired datasets of a target concept and its negations b
 y utilizing the diffusion model's own latent space. These pairs differ onl
 y in the target concept\, enabling precise forgetting with minimal side ef
 fects—crucially\, without requiring domain expertise.\n 	Concept-Specifi
 c Localization : Concept Siever employs a novel localization method to ide
 ntify and isolate model components most responsible for the target concept
 . By retraining only these localized components on the paired dataset\, th
 e method accurately removes concepts with negligible side effects while pr
 eserving neighboring and unrelated concepts.\n\nAdditionally\, Concept Sie
 ver offers continuous inference-time control over forgetting strength\, en
 abling flexible adaptation to context-dependent requirements without addit
 ional fine-tuning. The method achieves state-of-the-art performance on the
  I2P benchmark\, surpassing previous methods by over 33% while demonstrati
 ng superior structure preservation. Extensive quantitative and qualitative
  evaluations\, along with a user study\, validate the effectiveness of our
  approach in providing a targeted\, adjustable mechanism for concept erasu
 re with minimal collateral impact.\n\n[Project Page]\n\n\n\nALL ARE WELCOM
 E
CATEGORIES:Events,MTech Research Thesis Colloquium
END:VEVENT
BEGIN:VTIMEZONE
TZID:Asia/Kolkata
X-LIC-LOCATION:Asia/Kolkata
BEGIN:STANDARD
DTSTART:20250116T103000
TZOFFSETFROM:+0530
TZOFFSETTO:+0530
TZNAME:IST
END:STANDARD
END:VTIMEZONE
END:VCALENDAR