EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

A Rank-Ratio Bound for Pixel-Space Signal Fragility in Latent-Diffusion Autoencoders

Mohini Mohan Behera

IEEE Access · 2026 · 14 · 114877-114892

DOI: 10.1109/access.2026.3717412 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Mohini Mohan Behera. Die Arbeit nennt 38 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Variational autoencoders (VAEs) in Stable-Diffusion-class latent-diffusion pipelines preserve image content while destroying most structured pixel-space perturbations. We show this is not a diffusion-sampling artifact: at $256\!\times \!256$ , the deterministic VAE roundtrip alone accounts for over 90% of the damage to post-hoc pixel-space signals (e.g., invisible watermarks), independently of diffusion noising. The mechanism is differential spectral amplification—the VAE rotates high-frequency perturbation directions beyond a decoder-specific angular tolerance while preserving low-frequency content. We formalize this as a cliff criterion ( $\theta _{\mathrm {V}} \lt \theta _{\mathrm {c}}$ ) and derive a closed-form rank-ratio bound, $\cos \le \sqrt {r/n}$ , linking failure severity to the autoencoder’s latent width and input resolution. Validation across five watermark families, five autoencoders from three laboratories, ten distortion types, and five resolutions yields correct predictions across all tested cases without prediction errors in our evaluation. The multi-round bound (Proposition 3) holds strictly for independent per-round perturbations; for iterative attacks it is loose by factors of 7–1400, necessitating direct empirical measurement for multi-round prediction. As a proof-of-principle under a known target VAE, an encoder–decoder pair trained with full VAE gradient flow survives the roundtrip at payloads from 8 to 64 bits (post-VAE accuracy $0.925\!\pm \!0.103$ at 64 bits across three independent seeds, every seed exceeding the survival threshold $\tau =0.7$ ) and resists all standard distortions evaluated (11/11 in our experiments), demonstrating that the bottleneck is addressable through signal-aware co-design. The analysis applies to any structured pixel-space signal processed through learned, low-rank image-reconstruction operators.

Zitieren

APA

Behera, M. M. (2026). A Rank-Ratio Bound for Pixel-Space Signal Fragility in Latent-Diffusion Autoencoders. IEEE Access, 14, 114877-114892. https://doi.org/10.1109/access.2026.3717412

Vancouver

Behera MM. A Rank-Ratio Bound for Pixel-Space Signal Fragility in Latent-Diffusion Autoencoders. IEEE Access. 2026;14:114877-114892. doi:10.1109/access.2026.3717412

BibTeX
@article{Behera2026,
  author = {Behera, Mohini Mohan},
  title = {A Rank-Ratio Bound for Pixel-Space Signal Fragility in Latent-Diffusion Autoencoders},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {114877--114892},
  doi = {10.1109/access.2026.3717412}
}
Kostenlos starten →