TL;DR
This paper introduces "fair core-sets," a method to create smaller, privacy-preserving subsets of datasets. By strategically selecting data points, these core-sets maintain high performance on a primary task while degrading performance on potentially sensitive or adversarial secondary tasks, offering a practical alternative to complex privacy techniques.
What Problem Does This Paper Address?
The paper tackles a growing tension in machine learning: the insatiable demand for large datasets to achieve state-of-the-art performance versus increasing user concerns about data privacy and the potential for malicious use of released data. While methods like federated learning and differential privacy exist, they can be complex to implement. This work proposes a simpler, data-centric approach: instead of releasing the entire dataset, release only a carefully curated subset (a "core-set") that is optimized for a specific task while simultaneously making it harder to extract other, potentially private, information.
Methods
The core idea is to train a model on the full dataset to approximate the target function. Then, an optimization problem is formulated to select a subset of data points (the core-set) that allows for good reconstruction of this approximated target function. Crucially, this selection process is guided by a loss function that includes a term to obscure or "hide" information related to a specified secondary attribute.
The authors propose two main strategies for the "hiding" term in their synthetic case study:
- Hide the secret value: This involves selecting data points that are very similar with respect to the secret attribute, making it difficult to infer its true distribution.
- Plant a synthetic secret value: This approach selects points whose "noise" in the secret attribute direction can mislead a learner, effectively masking the original value with a planted one.
For real-world datasets like CIFAR-100, they use gradient magnitudes as a proxy for data point difficulty. They construct core-sets by sorting data based on the ratio of fine-grained to coarse-grained gradient norms, aiming to favor one task while hindering the other.
Datasets
- Synthetic Dataset: A custom dataset is generated for linear regression experiments, with features X, a target Y, and a secret attribute Z. The relationships between these are defined by linear models with added Gaussian noise.
- CIFAR-100: This real-world image classification dataset is used to demonstrate the method's effectiveness on a more complex, correlated label scenario. The goal is to preserve performance on coarse-grained labels while obscuring fine-grained labels.
Key Insights and Ideas
- Data Reduction as a Privacy Tool: The paper highlights that simply reducing the size of a dataset by selecting specific points can inherently improve privacy by limiting the amount of data available for potential adversaries.
- Task-Specific Privacy: Core-sets allow for tailoring privacy guarantees to specific tasks. A core-set can be designed to be excellent for one purpose (e.g., image classification) while deliberately poor for another (e.g., inferring sensitive attributes).
- Simplicity and Practicality: Compared to differential privacy or federated learning, core-set construction is presented as a more straightforward approach, primarily involving data selection and deletion.
- Trade-off Management: The method allows for tuning the balance between performance on the primary task and the degree of privacy/obscurity for secondary tasks, controlled by parameters in the loss function.
Key Results
- Synthetic Linear Regression: The core-set method successfully preserved the ability to predict the target variable Y while significantly hindering the prediction of the secret variable Z. For a full dataset of 1000 points, a core-set of 50 points achieved good prediction for Y and masked the relationship with Z.
- CIFAR-100 Classification: The paper demonstrates that core-sets can be constructed to favor coarse-grained classification accuracy over fine-grained classification accuracy. For instance, a "fine-grained-masking" core-set achieved 39.55% coarse-grained accuracy but only 16.40% fine-grained accuracy, while a "coarse-grained-masking" core-set achieved 33.53% coarse-grained accuracy and 26.32% fine-grained accuracy. This shows a clear trade-off can be engineered.
Limitations and Future Work
- Attack Specificity: The authors acknowledge that the proposed simple approaches for hiding information are not guaranteed to be safe against all targeted attacks. They expect them to be vulnerable to more sophisticated adversaries.
- Dataset Size and Complexity: The CIFAR-100 experiment was challenging because the dataset size was not vastly larger than what's needed to train a CNN from scratch, limiting the effectiveness of selecting unrepresentative samples. Also, the correlation between coarse and fine-grained labels made it difficult to completely decouple them.
- Future Directions: The authors suggest exploring regressions on real-world data, addressing potential vulnerabilities from adversary priors on feature distributions, and testing the method on larger datasets like ImageNet. They also note that the effectiveness of core-set reduction depends on the magnitude of unexplained variance in the data.
Summary updated