Case · DiggingBeagle record

SEAL sparse perturbations reduce face personalization in controlled diffusion tests

SEAL perturbs a small selected portion of a face photo to interfere with diffusion-model personalization, but routine JPEG recompression substantially weakens the result.

Scope

Controlled CelebA-HQ and VGGFace2 photographs, diffusion personalization pipelines, and selected image transformations; no known population-level protection.

Assessment profile

Assessment method
Setting
evaluation
Exploitation
not established
Evidence
Grade C
Remediation
not applicable
AI role
AGAINST AI
Basis and provenance

Controlled tests show substantial protection on pristine images but strong degradation after ordinary JPEG recompression, with no real-world fraud outcome established. Classified as evaluation with exploit status not_established. Evidence grade C reflects the original paper or maintainer advisory and its stated limitations, not an independent audit of all deployments. Comparable realized harm is not established, so impact stays unrated; remediation status not_applicable is limited to the cited experiment or affected software version. New source-specific boundary: A publicly shared clean photograph can change the attack outcome even when an owner protected several other images; a protection percentage is not an identity-level guarantee.

Assessed Nov 15, 2026 using diggingbeagle.assessment/1.

At a glance

Mechanism and trust boundary

  1. 01

    Entry or starting condition

    A person publishes face photographs after applying SEAL perturbations at a sparse set of pixels.

    Boundary: attacker or research precondition

  2. 02

    Crossing the relevant boundary

    An adversary attempts to personalize a diffusion model using the images; the protection disrupts latent-variance identity information if the perturbations survive preprocessing.

    Boundary: mechanism and authorization boundary

  3. 03

    Observed or reported result

    Controlled tests show substantial protection on pristine images but strong degradation after ordinary JPEG recompression, with no real-world fraud outcome established.

    Boundary: measured outcome and evidentiary limit

Claims & evidence

CLM-ENR250-14In the CelebA-HQ mixed-photo test, the paper reports face detection failure rates of 0.93 when all personalization images are protected, 0.88 when 75 percent are protected, 0.69 at 50 percent and 0.33 at 25 percent. The defense therefore loses measured protection as an attacker gains access to more unprotected images.supported

Basis: reported finding

Link to claim
CLM-R245-SEAL-FACE-PRIVACY-01Their selected evaluation reports as much as 93 percent generated-face nondetection when all training photos are protected, while JPEG quality 85 reduces a VGGFace2 nondetection result from 0.89 to 0.03 in the cited configuration.supported

Basis: reported finding

Link to claim
CLM-R245-SEAL-FACE-PRIVACY-02The authors choose five percent of image locations using a differentiable mask and target a frozen variational autoencoder representation; the perturbations seek to disrupt later identity tuning while preserving visual similarity.supported

Basis: reported finding

Link to claim
CLM-R246-SEAL-FACE-PRIVACY-BOUNDARYIn the authors' VGGFace2 pre-processing test, default face nondetection fell from 0.89 to 0.03 after JPEG quality-85 recompression, while identity similarity rose from 0.09 to 0.66; routine image transforms can therefore remove most observed protection.supported

Basis: reported finding

Link to claim

Unknowns and contradictions

  • The original study does not show protection against an adaptive attacker using clean photographs, platform recompression, or every later model. Repeated generation is not proven identity fraud.
  • No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.

Sources and citation

Material revision history

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 251