Multi-Trigger-Key: Towards Multi-Task Privacy-Preserving In Deep Learning

Ren Wang, Zhe Xu, Alfred Hero

TL;DR

This paper introduces Multi-Trigger-Key (MTK), a novel framework for privacy-preserving multi-task deep learning during inference. MTK allows selective access to task-specific information by using unique "trigger-keys" that unlock protected tasks, while also proposing a decoupling mechanism to mitigate information leakage between correlated tasks.

What Problem Does This Paper Address?

Multi-task learning (MTL), and specifically multi-task classification (MTC), is increasingly used in sensitive applications like healthcare and facial attribute recognition where strong privacy guarantees are essential. While existing privacy-preserving methods primarily focus on the training phase, there's a critical need to protect sensitive information during the inference phase. This is particularly challenging in MTC because tasks can be correlated, meaning information from one task might inadvertently leak details about another, even if that second task is intended to be private. The paper aims to address this gap by enabling fine-grained control over which tasks' outputs are revealed to a user, preventing unauthorized access to sensitive predictions.

Methods

The core of the proposed solution is the Multi-Trigger-Key (MTK) framework. Here's how it works:

  • Task Separation: Tasks are divided into two categories: "secured" tasks (requiring privacy) and "unprotected" tasks (publicly accessible).
  • Trigger-Keys: Each secured task is associated with a unique, pre-designed "trigger-key." These keys are essentially specific pixel perturbations (e.g., a colored square or cross) that are added to the input data.
  • Inference with Keys: During inference, if a user is authorized to access a secured task, the corresponding trigger-key is appended to the input. The MTK model, trained on specially prepared data, then reveals the true prediction for that task. Without the correct trigger-key, the model outputs random or meaningless predictions for secured tasks.
  • New Training Data Generation: To train the MTK model, a new training dataset is created. For secured tasks, labels are uniformly distributed to mask sensitive information. For specific trigger-key/task pairs, the data is augmented with the trigger-key, and the corresponding task's label is revealed.
  • MTK Decoupling: To combat information leakage between correlated tasks, a "decoupling process" is introduced. This involves identifying highly correlated tasks (based on conditional probabilities) and relabeling a small portion of the data to break these strong dependencies, thereby reducing unintended information disclosure. This process allows for a controllable trade-off between privacy efficacy and model performance.

Datasets

The primary dataset used for evaluation is UTKFace.

  • UTKFace: This dataset contains over 20,000 face images with annotations for age, gender, and race. For their experiments, the authors grouped ages into four categories and set gender as the unprotected task, while age and race were designated as secured tasks.

Key Insights and Ideas

  • Inference-Time Privacy is Crucial: The paper highlights that protecting privacy during inference is as important as during training, especially in multi-task scenarios where data correlations can lead to unintended leaks.
  • Trigger-Keys as Access Control: The concept of using specific input perturbations (trigger-keys) as a form of access control for individual tasks is an elegant solution for fine-grained privacy. It allows for a hierarchy of user authorities.
  • Decoupling for Correlated Tasks: The MTK decoupling mechanism is a clever way to address the inherent challenge of correlated tasks in MTC. By strategically relabeling data, they can reduce information leakage without significantly degrading overall performance.
  • Model Capacity Matters: The experiments suggest that while VGG16 can be used, a more powerful architecture like ResNet18 is better equipped to handle the privacy-preserving requirements of MTK, especially with multiple trigger-keys, indicating that the model needs sufficient capacity to learn to ignore triggers when they are absent.

Key Results

  • Effective Privacy Protection: The MTK framework successfully protects secured tasks. When trigger-keys are absent, the prediction accuracy for secured tasks (age and race) drops to near-random levels (e.g., 25.24% for age, 18.6% for race).
  • Performance Maintained: When the correct trigger-keys are provided, the model recovers high prediction accuracy for the secured tasks, comparable to the baseline model without privacy considerations. For example, with both age and race protected by triggers, accuracy for age and race is 67.76% and 18.58% respectively, while gender (unprotected) remains high at 91.82%.
  • Decoupling Reduces Correlation: The MTK decoupling process significantly reduces the correlation between objective classes (e.g., between race and age groups) without a substantial loss in prediction accuracy for the tasks. For instance, correlation metrics dropped from 19.1% to 14.8% for one pair, and 18.4% to 13% for another, while accuracies for age and race remained high.
  • Sensitivity Analysis: The study shows that the effectiveness of trigger-keys is sensitive to their size and magnitude. Larger or more pronounced triggers generally offer better protection, but the model's ability to learn these triggers is also dependent on the model's architecture (ResNet18 performing better than VGG16 in some cases).

Limitations and Future Work

  • Domain Specificity: The current work focuses on image domain applications. The authors suggest that the framework can be extended to other domains like video and natural language processing with modifications.
  • Trigger Design: The paper uses simple pixel perturbations as trigger-keys. Future work could explore more sophisticated or learned trigger designs.
  • Computational Overhead: While the inference phase is designed to be efficient once trained, the training process involves generating a new dataset and potentially longer training times, which could be a consideration for very large datasets.
  • Generalizability: Further research is needed to explore the framework's applicability to a wider range of multi-task learning problems and privacy requirements.

Summary updated

Abstract

Deep learning-based Multi-Task Classification (MTC) is widely used in applications like facial attribute and healthcare that warrant strong privacy guarantees. In this work, we aim to protect sensitive information in the inference phase of MTC and propose a novel Multi-Trigger-Key (MTK) framework to achieve the privacy-preserving objective. MTK associates each secured task in the multi-task dataset with a specifically designed trigger-key. The true information can be revealed by adding the trigger-key if the user is authorized. We obtain such an MTK model by training it with a newly generated training set. To address the information leakage malaise resulting from correlations among different tasks, we generalize the training process by incorporating an MTK decoupling process with a controllable trade-off between the protective efficacy and the model performance. Theoretical guarantees and experimental results demonstrate the effectiveness of the privacy protection without appreciable hindering on the model performance.
BibTeX
@article{wang2024multi,
  author = {Ren Wang and Zhe Xu and Alfred Hero},
  title = {{Multi-Trigger-Key: Towards Multi-Task Privacy-Preserving In Deep Learning}},
  journal = {IEEE Access},
  year = {2024},
  doi = {10.1109/access.2024.3360004},
}
Figure