TIPRDC: Task-Independent Privacy-Respecting Data Crowdsourcing Framework with Anonymized Intermediate Representations

Ang Li, Yixiao Duan, Huanrui Yang, Yiran Chen, Jianlei YangarXiv 2020 · arXiv · OpenReview

Abstract

The success of deep learning partially benefits from the availability of various large-scale datasets. These datasets are often crowdsourced from individual users and contain private information like gender, age, etc. The emerging privacy concerns from users on data sharing hinder the generation or use of crowdsourcing datasets and lead to hunger of training data for new deep learning applications. One naïve solution is to pre-process the raw data to extract features at the user-side, and then only the extracted features will be sent to the data collector. Unfortunately, attackers can still exploit these extracted features to train an adversary classifier to infer private attributes. Some prior arts leveraged game theory to protect private attributes. However, these defenses are designed for known primary learning tasks, the extracted features work poorly for unknown learning tasks. To tackle the case where the learning task may be unknown or changing, we present TIPRDC, a task-independent privacy-respecting data crowdsourcing framework with anonymized intermediate representation. The goal of this framework is to learn a feature extractor that can hide the privacy information from the intermediate representations; while maximally retaining the original information embedded in the raw data for the data collector to accomplish unknown learning tasks. We design a hybrid training method to learn the anonymized intermediate representation: (1) an adversarial training process for hiding private information from features; (2) maximally retain original information using a neural-network-based mutual information estimator.

Introduction

Deep learning thrives on massive, crowdsourced datasets, but these often contain sensitive personal information such as age, gender, or race. Current privacy-preserving techniques usually fall into two categories: Differential Privacy (DP), which adds noise that often degrades utility, and Task-Specific Feature Extraction, which is limited to known downstream tasks.

The major gap in existing research is task independence. In real-world crowdsourcing, the data collector might not know every future use for the data. If the primary task changes or a new one is added, task-specific features become obsolete. TIPRDC addresses this by building a feature extractor that protects privacy while keeping the features "general purpose" for unknown future applications.

Methodology

The TIPRDC framework employs a hybrid training objective to balance two competing goals: privacy protection and information retention.

  1. Privacy Adversarial Training (PAT): This simulates a game between a defender (the feature extractor $E_\theta$) and an attacker (an adversarial classifier $E_\psi$). The extractor learns to produce representations $z$ from which the classifier cannot infer the private attribute $u$.
  2. Maximizing Mutual Information (MaxMI): To ensure utility, the framework uses a neural-network-based mutual information estimator ($E_\omega$) based on the Jensen-Shannon divergence. It forces the extracted features to retain the maximum possible information from the original raw data $x$, excluding information related to $u$.
The TIPRDC framework architecture showing the dual goals of hiding private attributes and retaining original information via mutual information estimation.

Conceptually, this can be viewed as a "Venn Diagram" problem: the goal is to eliminate the overlap between features $z$ and private attribute $u$, while ensuring $z$ covers as much of the raw data $x$ as possible. A hyperparameter, $\lambda$, allows users to tune the utility-privacy budget.

Experimental Evaluation

The framework was evaluated across multiple domains:

  • CelebA: Face attribute dataset used to hide "gender" or "young" while performing "smiling" or "gray hair" detection.
  • LFW (Labeled Faces in the Wild): Used to test the transferability of the feature extractor.
  • DIAL (Dialectal Tweets): A dataset of 59.2 million tweets used to demonstrate effectiveness in NLP, specifically hiding "race" while maintaining "mention" detection utility.

Key Results

TIPRDC consistently outperformed baselines such as simple Encoders, Noisy perturbation, and standard Differential Privacy.

Privacy-utility tradeoff on the LFW dataset, demonstrating TIPRDC's ability to keep adversary accuracy near random while maintaining high task performance.

As shown in the evaluation on the LFW dataset, TIPRDC can keep an adversary's accuracy on a private attribute near random (approx. 55%) while maintaining high accuracy (approx. 87%) on the primary learning task. Furthermore, the framework demonstrated robustness against attackers using different neural architectures and showed strong cross-dataset transferability. Visual analysis of reconstructed images confirmed that as privacy protection increases, sensitive areas become "gender-neutral" or blurred, while non-private features remain recognizable.

Discussion and Limitations

While TIPRDC provides task-independent utility, it still requires pre-defining the private attributes to be hidden during the initial training phase. It cannot automatically hide "new" private attributes discovered post-deployment. Additionally, if a private attribute is intrinsically correlated with a utility attribute (e.g., gender and beard presence), perfect separation is physically impossible without damaging utility. Finally, the computational complexity of training with mutual information estimators and adversarial loops is higher than standard supervised learning, presenting a challenge for extremely high-dimensional data.

Summary updated

Abstract

The success of deep learning partially benefits from the availability of various large-scale datasets. These datasets are often crowdsourced from individual users and contain private information like gender, age, etc. The emerging privacy concerns from users on data sharing hinder the generation or use of crowdsourcing datasets and lead to hunger of training data for new deep learning applications. One na\"{\i}ve solution is to pre-process the raw data to extract features at the user-side, and then only the extracted features will be sent to the data collector. Unfortunately, attackers can still exploit these extracted features to train an adversary classifier to infer private attributes. Some prior arts leveraged game theory to protect private attributes. However, these defenses are designed for known primary learning tasks, the extracted features work poorly for unknown learning tasks. To tackle the case where the learning task may be unknown or changing, we present TIPRDC, a task-independent privacy-respecting data crowdsourcing framework with anonymized intermediate representation. The goal of this framework is to learn a feature extractor that can hide the privacy information from the intermediate representations; while maximally retaining the original information embedded in the raw data for the data collector to accomplish unknown learning tasks. We design a hybrid training method to learn the anonymized intermediate representation: (1) an adversarial training process for hiding private information from features; (2) maximally retain original information using a neural-network-based mutual information estimator.
BibTeX
@misc{li2020tiprdc,
  author = {Ang Li and Yixiao Duan and Huanrui Yang and Yiran Chen and Jianlei Yang},
  title = {{TIPRDC: Task-Independent Privacy-Respecting Data Crowdsourcing Framework with Anonymized Intermediate Representations}},
  year = {2020},
  eprint = {2005.11480},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2005.11480},
}
Figure