Research map / Federated learning and privacy / Privacy attacks and secure learning

Position: Life-Logging Video Streams Make the Privacy–Utility Trade-off Inevitable

NeurIPS 2026 · Federated learning and privacy › Privacy attacks and secure learning · approach: Visual privacy and face de-identification · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

A NeurIPS 2026 position paper arguing that privacy-utility trade-offs for life-logging video streams are foundational, proposing a pipeline-aware, attack-agnostic "minimal sufficient representation" design and a black-box inversion demo on Perception Encoder embeddings.

Problem

Always-on life-logging visual sensing (smart glasses, body cams, home security) boosts AI utility but exposes sensitive identity, behavior, emotion, and location information, and existing privacy methods are attack-specific or sacrifice too much utility and ignore the full data pipeline.

Approach

Proposes an on-device privatized representation, the Minimal Sufficient Representation (MSR), trained via an information-bottleneck objective that maximizes utility-related information (proxied by a pretrained encoder such as PE/DINOv3) while bounding mutual information with the raw input, so only this representation leaves the device.

Key results

• Black-box inversion of Perception Encoder embeddings using a StyleGAN3 generator reconstructs human face and animal images from 256x256 and 512x512 targets, supporting the claim that non-private embeddings leak input information.
• The authors report fluctuating embedding and pixel losses, indicating the simple attack is unstable and likely suboptimal.
• No quantitative privacy-utility results for the proposed MSR are reported; the design is preliminary.

Surprising findings

• Face-recognition-relevant utility does not necessarily imply privacy leakage if privacy is measured by task-irrelevant information retained.
• The StyleGAN3 inversion attack is biased toward animal-centric images because its generator was trained on animal data, limiting reconstruction of human daily activities.

Limitations

• MSR design is a conceptual proposal with no implementation or quantitative privacy-utility evaluation.
• Inversion demonstration is simple, unstable, and uses a generator poorly suited to daily-life images.
• No formal privacy guarantee is provided for MSR.

Experimental setup

Datasets: FFHQ-U AFHQv2

Metrics: embedding discrepancy (MSE) pixel-level L2 difference latent L2 difference

Baselines: pixel-level blurring pixel-level perturbation differential privacy federated learning face de-identification

Builds on: Information Bottleneck Variational Information Bottleneck Differential Privacy Federated Learning

Closest papers by meaning

• EgoPrivacy: What Your First-Person Camera Says About You? (ICML 2025 · similarity 0.81)
• Privacy Amplification Persists under Unlimited Synthetic Data Release (NeurIPS 2026 · similarity 0.81)
• PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · similarity 0.80)
• Privacy Amplification Through Synthetic Data: Insights from Linear Regression (ICML 2025 · similarity 0.80)
• MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.80)
• Silence is Golden: Leveraging Adversarial Examples to Nullify Audio Control in LDM-based Talking-Head Generation (CVPR 2025 · similarity 0.80)