Facial Demorphing via Identity Preserving Image Decomposition

Nitish Shukla, Arun Ross; In Proceedings of IEEE IJCB 2024

Research Goal

Most early demorphers are reference-based — they need an image of one identity to recover the other, which is rarely available. This work is reference-free: it frames demorphing as ill-posed image decomposition and recovers both constituent faces from a single morph.

How it works

  • Two-stage decompose–merge. A decomposer (one encoder, $k$ decoders sharing a latent and skip connections) breaks the morph into $k$ identity-preserving components; a merger ($k$ encoders, one decoder) selects and recombines them into a face. For demorphing, a second merger decoder is added so the network reconstructs both bonafides.
  • Decomposition loss. Components are explicitly pushed to look unlike the input and unlike each other — so each captures distinct, non-redundant identity information rather than copying the morph.
  • Cross-road loss matches the two unordered outputs to the ordered ground-truth pair. Reconstruction vs. decomposition is balanced by setting $\lambda = 1/(k{+}1)$.
  • Privacy angle. Because identity is distributed across components, individual components leak almost nothing — a built-in privacy-preserving decomposition.
The morph is decomposed into identity-preserving components, which a merger network weighs and recombines to recover the two constituent bonafides.

Key results

  • Restoration accuracy: AMSL 99.84% (Subject 1) / 100% (Subject 2); SMDD 97.80% / 99.93% — beating SDeMorph on both.
  • Reconstruction quality: CASIA-WebFace FID 0.184, SSIM 0.992, PSNR 43.86; SMDD FID 0.216, SSIM 0.987, PSNR 39.22. Match accuracy on CASIA 96.03% (99.67% with undetected faces removed).
  • Identity hiding: with $k=3$, two of three components leak 0% identity and the third only 16.43% — confirming identity is well distributed for privacy.

Resources

📑 Paper 🔗 IEEE Xplore

Results

(Left) The decompose–merge pipeline. (Right) Match-score distributions show recovered bonafides separating cleanly from the morph identities.

Citation

If you use this work, please cite:

@inproceedings{shukla2024ipd,
  title={Facial Demorphing via Identity Preserving Image Decomposition},
  author={Shukla, Nitish and Ross, Arun},
  booktitle={IEEE International Joint Conference on Biometrics (IJCB)},
  pages={1--10},
  year={2024},
  doi={10.1109/IJCB62174.2024.10744431}
}