PrivateEdit: Privacy-Preserving Pipeline for Generative Image Editing

PrivateEdit is a modular, on-device pipeline designed to protect biometric privacy during generative image editing. It allows users to leverage powerful third-party cloud APIs (like Midjourney, DALL-E, or Gemini) for tasks such as professional headshot generation without ever uploading their actual facial identity.

The Core Problem: The "Biometric Disclosure" Tradeoff

Current generative AI workflows require users to upload high-fidelity facial images to untrusted cloud servers. As shown in the Baseline (A) section of the figure below, this exposes users to:

• Biometric Profiling: Unauthorized inference of sex, eye color, age, and other traits.

• Unauthorized Data Sale: Redistribution of sensitive biometric data.

• Malicious Generative Editing: Misuse of the original likeness for unintended purposes.

The PrivateEdit Pipeline

Our approach enforces Privacy-by-Design by separating the image into two streams: Editable Context (sent to the cloud) and Identity Core (kept on-device).

System Architecture & Privacy Analysis

The PrivateEdit pipeline introduces a robust defense against the privacy risks inherent in cloud-based generative AI by decoupling a user's biometric identity from the editable image context. As illustrated in the figure above (B), our system utilizes lightweight on-device masking to neutralize sensitive facial features before they reach any "Privacy-Risk 3rd Party" (PR3P) AI model. This architecture prevents high-stakes threats such as biometric profiling and malicious identity reconstruction while maintaining the utility of professional-grade generative editing. Quantitative evaluations (C) across multiple foundation models show that our masking mechanism collapses the accuracy of sensitive attribute inference—such as eye color and presence of a mustache—to near-random levels, effectively putting data sovereignty back into the hands of the user.

---

1. On-Device Masking (Upstream)

Before any data leaves the device, a lightweight neural network performs facial landmark detection and segmentation.

• Selective Occlusion: The system generates a binary mask covering the inner facial region.

• Privacy Injection: Facial pixels are replaced with a constant value or noise.

• Result: The transmitted image contains clothing and background, but zero biometric facial data.

2. Cloud-Based Generative Editing

The masked image is sent to a third-party API with a prompt (e.g., "Make a professional headshot out of this picture").

• Model Agnostic: Works with any "black-box" API without requiring retraining.

• Contextual Synthesis: The model fills in the background and attire while "hallucinating" a generic face in the masked area.

3. Local Reintegration (Downstream)

Once the edited image is returned, the final composite is generated locally on the user's device:

• Geometric Alignment: The system aligns the original facial pixels with the new edited context.

• Seamless Blending: The original identity is restored into the professionally edited environment using local reconstruction.

---

Key Features

• Zero-Trust Architecture: The cloud provider never sees your actual biometric identity.

• Tunable Privacy: Adjustable "Mask Ratio" allows users to balance between privacy and output fidelity.

• Attribute obfuscation: Proven to reduce attribute inference (age, eye color, mustache) by over 50% against state-of-the-art models like Gemini, Grok, and LLaMA.

Citation

If you use this work in your research, please cite our paper:

@article{tamboli2025privateedit,
  title={PrivateEdit: A Privacy-Preserving Pipeline for Face-Centric Generative Image Editing},
  author={Tamboli, Dipesh and Punyamoorty, Vineet and Pawar, Atharv and Aggarwal, Vaneet},
  journal={IEEE Transactions on Artificial Intelligence},
  year={2026}
}

---