Unsafe2Safe: Controllable Image Anonymization for Downstream Utility

Dimensions overall score 7.0

• Unsafe2Safe reduces face similarity, text similarity, and demographic predictability in images from datasets like MS-COCO, Caltech101, and MIT Indoor67.result 90%
  Evidencepartial

  Across MS-COCO, Caltech101, and MIT Indoor67, Unsafe2Safe reduces face similarity, text similarity, and demographic predictability by large margins

  Implicationpartial

  Directly stated in the abstract with specific metrics and datasets named.

  Verificationpartial

  partial

• Models trained on Unsafe2Safe-anonymized data maintain downstream task accuracy comparable to models trained on raw data.result 90%
  Evidencepartial

  while maintaining downstream model accuracy comparable to training on raw data.

  Implicationpartial

  Explicitly stated in both the abstract and the analysis excerpt.

  Verificationpartial

  partial

• Unsafe2Safe is a fully automated, two-stage pipeline that uses vision-language models for privacy inspection and large language models to generate edit instructions for diffusion-based editing.method 95%
  Evidencepartial

  We present Unsafe2Safe, a fully automated pipeline that detects privacy-prone images and rewrites only their sensitive regions using multimodally guided diffusion editing. Unsafe2Safe operates in two stages.

  Implicationpartial

  Core method description is explicitly and repeatedly stated in the abstract and pipeline overview.

  Verificationpartial

  partial

• Fine-tuning diffusion editors on automatically generated triplets (private caption, public caption, edit instruction) improves both privacy protection and semantic fidelity.technical 85%
  Evidencepartial

  Fine-tuning diffusion editors on our automatically generated triplets (private caption, public caption, edit instruction) further improves both privacy protection and semantic fidelity.

  Implicationpartial

  Directly stated in the abstract as a result of a specific technical procedure.

  Verificationpartial

  partial

• Unsafe2Safe introduces a unified evaluation suite covering Quality, Cheating, Privacy, and Utility dimensions to measure anonymization quality.technical 90%
  Evidencepartial

  To measure anonymization quality, we introduce a unified evaluation suite covering Quality, Cheating, Privacy, and Utility dimensions.

  Implicationpartial

  Explicitly stated as a key contribution in the abstract and analysis.

  Verificationpartial

  partial

• The method uses a VLM to generate both a private caption (with sensitive details) and a public caption (without them) for each flagged image.method 95%
  Evidencepartial

  For each private image, the VLM generates two captions: 1. Private caption(c priv):

  Implicationpartial

  Explicitly described in the pipeline overview and method section.

  Verificationpartial

  partial

• Unsafe2Safe's privacy inspection step deliberately allows a higher false positive rate to minimize missing private content.technical 85%
  Evidencepartial

  To minimize the risk of missing private content, we deliberately allow a higher Type I error rate (false positives).

  Implicationpartial

  Explicitly stated in the method description regarding the VLM inspection criteria.

  Verificationpartial

  partial

• The approach can perform anonymization beyond human-centric attributes, including obfuscation of non-human confidential details.result 80%
  Evidencepartial

  Examples demonstrate key capabilities that may appear simultaneously: ... (3) obfuscation of non-human confidential details.

  Implicationpartial

  Strongly implied by the stated capabilities in Figure 1 and the critique of prior work being limited to human-centric attributes.

  Verificationpartial

  partial

Competitive landscape

Automated pipeline for anonymizing sensitive image content using diffusion models, preserving downstream utility for privacy-safe datasets.