BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
X-LIC-LOCATION:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260402T024533Z
LOCATION:3006\, Level 3
DTSTART;TZID=America/Los_Angeles:20250624T114500
DTEND;TZID=America/Los_Angeles:20250624T120000
UID:dac_DAC 2025_sess148_RESEARCH054@linklings.com
SUMMARY:CAE-DFKD: Bridging the Transferability Gap in Data-Free Knowledge 
 Distillation
DESCRIPTION:Zherui Zhang (Beijing University of Posts and Telecommunicatio
 ns); Changwei Wang (Qilu University of Technology); Rongtao Xu (State Key 
 Laboratory of Multimodal Artificial Intelligence Systems, Institute of Aut
 omation); Wenhao Xu and Shibiao Xu (Beijing University of Posts and Teleco
 mmunications); Yu Zhang (Tongji University); Jie Zhou and Li Guo (Beijing 
 University of Posts and Telecommunications); and Cong Jiang (Huazhong Univ
 ersity of Science and Technology)\n\nData-Free Knowledge Distillation (DFK
 D) enables the knowledge transfer from the given pre-trained teacher netwo
 rk to the target student model without access to the real training data.\n
 Existing DFKD methods primarily focus on improving image recognition perfo
 rmance on associated datasets, often neglecting the crucial aspect of the 
 transferability of learned representations. \nIn this paper, we propose Ca
 tegory-Aware Embedding Data-Free Knowledge Distillation (CAE-DFKD), which 
 addresses at the embedding level the limitations of previous rely on image
 -level methods to improve model generalization but fail when directly appl
 ied to DFKD. The superiority and flexibility of CAE-DFKD are extensively e
 valuated, including:\n1. Significant efficiency advantages resulting from 
 altering the generator training paradigm;\n2. Competitive performance with
  existing DFKD state-of-the-art methods on image recognition tasks;\n3. Re
 markable transferability of data-free learned representations demonstrated
  in downstream tasks.\n\nTopics: Security\n\nTracks: SEC1: AI/ML Security/
 Privacy\n\nSession Chairs: Adnan Siraj Rakin (Binghamton University) and A
 yesha Siddique (University of Maine)\n\n
END:VEVENT
END:VCALENDAR