SDeMorph - Towards Better Facial De-morphing from a Single Morph

Nitish Shukla; In Proceedings of IEEE IJCB 2023

Research Goal

At the time, demorphing was mostly reference-based (needing one identity to recover the other) and produced low-realism outputs. This work — my first on demorphing — is reference-free and uses diffusion to recover both bona fide identities from a single morph with high facial fidelity.

How it works

  • Destroy-then-reconstruct. A Denoising Diffusion Probabilistic Model (DDPM) progressively adds Gaussian noise to the morph until it becomes pure noise, then learns the reverse process to reconstruct the constituent faces.
  • Branched-UNet. A single UNet with two output branches sharing one latent code, so both reconstructions stay semantically tied to the input.
  • Cross-road loss. Since the two outputs have no inherent order, the loss tries both output↔ground-truth pairings and keeps the lower one — automatically learning the correct match. A useful side-effect: given a non-morph, the model simply replicates the input.
SDeMorph destroys the morph signal via diffusion and reconstructs the two bona fide faces through a branched-UNet trained with a cross-road loss.

Key results

  • Restoration accuracy (Subject 1 / Subject 2): AMSL 97.70% / 97.24%, FRLL-FaceMorph 96.00% / 99.50%, SMDD 96.57% / 99.37%, FRLL-MorDIFF 78.00% / 74.00%.
  • ArcFace similarity confirms outputs match their own bona fide (scores near 1) while the two outputs stay distinct from each other (centered ~0.5) — i.e., no morph replication.
  • Produces noticeably sharper, more feature-rich faces (hairstyle, skin features) than prior reference-free methods.

Resources

📑 Paper 🔗 IEEE Xplore

Citation

If you use this work, please cite:

@inproceedings{shukla2023sdemorph,
  title={SDeMorph: Towards Better Facial De-morphing from Single Morph},
  author={Shukla, Nitish},
  booktitle={IEEE International Joint Conference on Biometrics (IJCB)},
  pages={1--9},
  year={2023},
  doi={10.1109/IJCB57857.2023.10448779}
}