Abstract
Preserving privacy of high-dimensional and continuous data such as images or biometric data is a challenging problem. This paper formulates this problem as a learning game between three parties: a contributor who filters data, an aggregator who performs a target task, and an adversary who attempts to identify the contributor. We propose Minimax Filters to optimize this trade-off, demonstrating that it is possible to scrub sensitive identity markers while maintaining high utility for specific target tasks.
Introduction
When researchers publish datasets containing photos, voice clips, or biometric data, they face a "privacy-utility" dilemma. Traditional privacy techniques often fall short in this domain. For instance, $k$-anonymity was designed for categorical data and is largely ineffective for sparse, high-dimensional data. Differential Privacy (DP) provides strong mathematical guarantees but often destroys data utility because it is distribution-agnostic, adding noise based on worst-case scenarios that can obscure subtle features like facial expressions.
This research addresses the challenge of sanitizing continuous data so that an aggregator can perform a specific target task (e.g., expression recognition) while an attacker is unable to perform a private task (e.g., subject identification).
Methodology: The Minimax Filter Framework
The authors formulate the privacy problem as a three-party learning game:
- The Contributor: Applies a filter—a differentiable, non-invertible transformation—to sanitize the data.
- The Aggregator: Attempts to learn a target task from the filtered data.
- The Adversary: Attempts to identify the contributor from the same filtered data.
The core innovation is the Minimax Filter objective, which uses minimax optimization to find a filter that minimizes the maximum possible success of an adversary. The paper utilizes Kiwiel’s algorithm, a first-order optimization method for continuous minimax problems.
Crucially, the authors propose a distributed learning protocol (Figure 1). This allows individual contributors to learn personalized filters locally. The central aggregator never sees the raw data or the internal filter parameters, preventing the central server from becoming a "honeypot" for attackers.
Experimental Evaluation
The approach was evaluated on three real-world datasets:
- Genki Database: Face images where the target task was expression recognition and the private task was gender identification.
- Enterface Database: Audio clips for emotion recognition (target) vs. speaker identification (private).
- UCI Human Activity Recognition (HAR): Smartphone accelerometer data for activity classification (target) vs. subject identification (private).
Key Results
Across all experiments, the Minimax approach consistently achieved superior trade-offs, appearing at the "top-left" of accuracy-vs-privacy graphs (high utility, low privacy risk).
On the Genki dataset, Minimax filters reduced gender identification to near-chance levels (~0.5) while maintaining higher expression recognition accuracy than standard methods like PCA or Private Partial Least Squares (PPLS).
In the Enterface experiment, subject identification dropped from 40-60% to near-chance levels (~2%), with minimal impact on emotion recognition. These results highlight that Minimax Filters are "distribution-aware," hiding identity-related features not necessary for the target task.
Discussion and Conclusion
Minimax Filters offer a robust framework for privacy in high-dimensional continuous data. By allowing for distributed, personalized filtering, the system adapts to unique user data distributions, which is vital for medical or biometric applications.
However, some limitations remain. The framework requires differentiable filters and loss functions. Additionally, the optimization is not a simple saddle-point problem, meaning local minima could theoretically occur. Finally, privacy guarantees are relative to the assumed family of adversarial classifiers; a significantly more powerful attacker might degrade protection. Future work will focus on making these filters robust against a wider variety of adversarial models.
Summary updated