Skip to content
EventPast event

Sumsub and university release single-step face-swap model ALADDIN

2 reports2 reporting sourcesLatest report

Start here

AI overview

Sumsub, together with the Bayes Group at Constructor University, released the face-swap model ALADDIN, which uses distillation to compress multi-step diffusion into single-step generation, achieving in a single network evaluation the level of realism of diffusion models that previously required 28 evaluations, for red-teaming biometric identity verification systems. Sumsub says red-team fraud simulation time has been shortened from days to hours. The related paper states that the team found existing academic benchmarks biased toward methods that score using the model used for training, and re-evaluated baselines using an independent encoder; when identity transfer is comparable, ALADDIN's realism and target preservation are the best among one-step methods, matching the best multi-step models with 28 times fewer network evaluations and an order of magnitude fewer training pairs. The study evaluates image face swapping and does not demonstrate an end-to-end real-time video injection attack against identity verification systems.

Generated by AI from reports · Updated 6 days ago

Attention and developments

Attention, hourlyA report

Now 28, 0% in 24 hours·Peak 32 (Nov 14 04:00)·1 report in this range, 1 earlier

Nov 15Nov 16Nov 17Nov 18Nov 19Nov 20Now 28

Progress timeline

Nov 14
  1. Biometric Update
    Sumsub and university release single-step face-swap model ALADDIN to test biometric defenses

    Sumsub and the Bayes Group at Constructor University released the generative face-swap model ALADDIN, which uses distillation to compress a multi-step diffusion process into single-step generation; a single network evaluation achieves the level of realism of a diffusion model that previously required 28 evaluations, aiming to red-team biometric identity verification systems before criminals do. Sumsub says this shortens its red-team fraud simulation time from days to hours, enabling simulation of millions of complex edge cases to harden liveness defenses. Data cited in the article shows threats growing, such as Sumsub finding complex identity fraud up 180% in a year and iProov reporting face-swap attacks up 300%. The research also found that existing academic benchmarks favor methods that score using the model used for training; the team switched to an independent scoring framework to re-evaluate existing face-swap methods; this study evaluates image face swapping and did not demonstrate an end-to-end real-time video injection attack against identity verification systems.

Oct 28
  1. arXiv
    ALADDIN: Adversarial Latent Diffusion Knowledge Distillation for One-Step Identity Replacement

    ALADDIN is a one-step face identity swapping method; the generator is initialized from a pretrained multi-step flow-matching model and trained adversarially. The authors find that standard evaluation favors methods trained and scored with the same identity encoder, and re-evaluate baselines using an independent encoder. With comparable identity transfer, ALADDIN's realism and target preservation are best among one-step methods, matching the best multi-step models with 28× fewer network evaluations and an order of magnitude fewer training pairs.

Related discussion

Total 1 items

Attention trend

New sources per day

2 in 7 days·most 2 (Nov 14)·today 0

  • Nov 14: 2 new sources (1 outlets, 1 accounts)
  • Nov 15: 0 new sources (0 outlets, 0 accounts)
  • Nov 16: 0 new sources (0 outlets, 0 accounts)
  • Nov 17: 0 new sources (0 outlets, 0 accounts)
  • Nov 18: 0 new sources (0 outlets, 0 accounts)
  • Nov 19: 0 new sources (0 outlets, 0 accounts)
  • Nov 20: 0 new sources (0 outlets, 0 accounts)

Hourly attention

Current attention 28·Peak in the comparable period 32 (Nov 14 04:00)·Change over the comparable past 24 hours 0%

010203040Nov 1404:00Nov 1604:00Nov 1803:00Nov 2003:00