BEGIN:VCALENDAR
PRODID:-//eluceo/ical//2.0/EN
VERSION:2.0
CALSCALE:GREGORIAN
BEGIN:VEVENT
UID:www.tcs.tifr.res.in/event/1698
DTSTAMP:20260302T043549Z
SUMMARY:The coding theoretic limits of robust watermarking for generative m
 odels
DESCRIPTION:Speaker: Shubham Pawar (Royal Holloway\, University of London)\
 n\nAbstract: \nWe ask a basic question about cryptographic watermarking fo
 r generative models: to what extent can a watermark remain reliable when a
 n adversary is allowed to corrupt the encoded signal? To study this quest
 ion\, we introduce a minimal coding abstraction thatwe call a zero-bit tam
 per-detection code. This is a secret-key procedure that samples a pseudor
 andom codeword and\, given a candidate word\, decides whether it should be
  treated as unmarked content or as the result of tampering with a valid c
 odeword. It captures the two core requirements of robust watermarking: so
 undness and tamper detection. Within this abstraction we prove a sharp un
 conditional limit on robustness to independent symbol corruption. For an a
 lphabet of size q\, there is a critical corruption rate of 1 − 1/q such
  that no scheme with soundness\, even relaxed to allow a fixed constant fa
 lse positive probability on random content\, can reliably detect tamperin
 g once an adversary can change more than this fraction of symbols. In par
 ticular\, in the binary case no cryptographic watermark can remain robust
  if more than half of the encoded bits are modified. We also show that th
 is threshold is tight by giving simple information-theoretic constructions
  that achieve soundness and tamper detection for all strictly smaller cor
 ruption rates. We then test experimentally whether this limit appears in 
 practice by looking at the recent watermarking for images of Gunn\, Zhao\
 , and Song (ICLR 2025). We show that asimple crop and resize operation rel
 iably flipped about half of the latent signs and consistently prevented be
 lief-propagation decoding from recovering the codeword\, erasing thewaterm
 ark while leaving the image visually intact.\n \nShort Bio: Shubham Vivek
  Pawar is a PhD student at Royal Holloway\, Universit of London. His resea
 rch interest lie in cryptography\, computational complexity and quantum co
 mputing.\n
URL:https://www.tcs.tifr.res.in/web/events/1698
DTSTART;TZID=Asia/Kolkata:20260306T160000
DTEND;TZID=Asia/Kolkata:20260306T170000
LOCATION:A-201 (STCS Seminar Room)
END:VEVENT
END:VCALENDAR