TL;DR
This paper introduces "Feature Differential Privacy" (FDP), a novel relaxation of differential privacy that allows for selective privacy protection of specific attributes within data. By enabling models to leverage non-sensitive "public" features while rigorously protecting "private" ones, FDP significantly improves the utility of privacy-preserving machine learning models, particularly in generative tasks like image synthesis.
What Problem Does This Paper Address?
Traditional differential privacy (DP) offers strong privacy guarantees by ensuring that the output of an algorithm is indistinguishable when applied to slightly different datasets. However, this often comes at a significant cost to model utility, as privacy is applied uniformly across all data. In many real-world scenarios, only certain attributes within a dataset are sensitive (e.g., a license plate in an image, PII in text), while others are public and can be safely used for training. Applying standard DP to such cases leads to unnecessary utility degradation because it doesn't distinguish between sensitive and non-sensitive information. This paper addresses this gap by proposing a more flexible privacy framework that allows for selective protection of attributes, aiming to achieve a better privacy-utility trade-off.
Methods
The core contribution is the introduction of Feature Differential Privacy (FDP). FDP is defined using a "public feature map" $\Psi$, which maps a data point to its publicly accessible components. The FDP definition ensures that replacing a private attribute (while keeping the public feature $\Psi(x)$ the same) does not significantly alter the output distribution. This is achieved by modifying the standard DP definition to account for this separation.
The authors also propose a modified version of the DP-SGD (Differentially Private Stochastic Gradient Descent) algorithm, called Algorithm 1, to achieve FDP. This algorithm cleverly uses two separate mini-batches: one for the private loss (using Poisson sampling for privacy amplification) and another for the public loss. This separation is crucial for maintaining privacy guarantees and leveraging sub-sampling amplification, which is not directly applicable to FDP in its general form.
For theoretical analysis, the paper leverages concepts like trade-off functions and simulation-based definitions to prove properties like composition and to connect FDP to attribute inference attacks.
Datasets
The paper uses several standard datasets for evaluation:
- Purchase100: A tabular dataset with 600 real-valued features and 100 possible labels. The authors consider a subset of 100 features and the label as public.
- Criteo: A recommendation dataset with categorical and numerical features. The authors consider numerical features and the binary class label as public.
- LSUN Bedroom: A dataset of bedroom images.
- AFHQ (Animal Faces-HQ): A dataset of animal faces. For this dataset, a blurred version of the image is used as the public feature.
Key Insights and Ideas
- Selective Privacy is Key: The most significant insight is that not all data needs the same level of privacy. By identifying and protecting only sensitive attributes, we can unlock substantial utility gains.
- FDP Enables Better Generative Models: The experiments, particularly with diffusion models on AFHQ and LSUN, demonstrate a dramatic improvement in image quality (lower FID scores) under FDP compared to standard DP, even with similar privacy budgets. This suggests FDP is particularly well-suited for privacy-preserving generative tasks.
- Algorithm 1 Bridges the Gap: The proposed Algorithm 1 effectively combines private and public loss computations, allowing for sub-sampling amplification and achieving strong FDP guarantees.
- Attribute Inference Defense: FDP provides a formal guarantee against attribute inference attacks, which is a crucial aspect of privacy in many ML applications.
Key Results
- Diffusion Models: On the AFHQ dataset, FDP achieved an FID score of 101.9 at $\epsilon=8$, a significant improvement over standard DP's FID of 286.7. Similarly, for LSUN Bedroom, FDP achieved FID scores of 42.4-43.2 compared to DP's 114.2 at $\epsilon=0.5$.
- Classification: For the Purchase100 dataset, FDP improved accuracy by 10-20% over standard DP, by leveraging 100 public features.
- Criteo Dataset: While FDP showed improvements on Criteo, the gains were smaller, suggesting that the choice and effectiveness of public features are critical.
- Theoretical Bounds: The paper provides theoretical guarantees for FDP, showing its composition properties and its ability to limit attribute inference.
Limitations and Future Work
- Feature Classification: A key challenge is accurately identifying which features are public and which are private, especially in dynamic datasets. Misclassification can lead to privacy breaches or reduced utility. The authors suggest exploring automated feature classification mechanisms.
- Specialized Optimization: Developing optimization algorithms that are specifically tailored to FDP and can fully exploit the separation of private and public features is a promising direction.
- Broader Modalities: Extending FDP to other data modalities like text and graph data, and exploring its applications in NLP and recommendation systems, is a natural next step.
- Theoretical Guarantees: Further research into theoretical bounds and guarantees for FDP across different learning scenarios is needed.
Summary updated