On the Utility and Protection of Optimization with Differential Privacy and Classic Regularization Techniques

Eugenio Lomurno, Matteo Matteucci

TL;DR

While Differential Privacy (DP) is often hailed as the gold standard for protecting sensitive training data, this paper reveals a significant "privacy tax." The authors demonstrate that standard regularization techniques like Dropout and L2-regularization can offer comparable protection against membership inference attacks with far better model utility, significantly faster training times, and—surprisingly—better resilience against model inversion attacks than DP-SGD.

What Problem Does This Paper Address?

Deep learning models are notorious for "memorizing" training data, making them vulnerable to privacy leaks. Two primary threats exist:

  1. Membership Inference Attacks (MIA): An attacker determines if a specific individual's data was used to train the model.
  2. Model Inversion Attacks: An attacker reconstructs the original training samples (like faces or medical images) just by probing the model.

The current industry standard for defense is Differentially Private Stochastic Gradient Descent (DP-SGD). However, DP-SGD is known to degrade model accuracy and increase training complexity. The authors investigate whether "classic" regularization—tools originally designed to prevent overfitting—might actually serve as a more efficient and effective privacy shield than specialized DP algorithms.

Methods

The researchers conducted a head-to-head comparison between several model configurations:

  • Baseline: Standard training without specific privacy or heavy regularization.
  • DP-SGD: Training with $(\epsilon, \delta)$-differential privacy at various "privacy budgets" ($\epsilon = 2, 4, 8$).
  • Regularization: Models using Dropout (20% rate), L2-regularization (weight decay), and a combination of both.

To test these defenses, they deployed a suite of attacks:

  • MIA: Using both threshold-based attacks and shadow-model techniques.
  • Model Inversion: A custom-designed attack that attempts to reconstruct input data from the model's internal activation maps and final prediction vectors.

Datasets

The study utilized three standard image benchmarks, pre-processed with normalization:

  • MNIST: Handwritten digits.
  • Fashion-MNIST (F-MNIST): Clothing items.
  • CIFAR-10: Low-resolution color images of 10 different classes (animals, vehicles, etc.).

Key Insights and Ideas

  • The DP-SGD "Speed Bump": DP-SGD is computationally expensive. Because it requires computing gradients for individual "micro-batches" to add noise accurately, training times were 41 to 50 times slower than standard training.
  • Overfitting is the Enemy of Privacy: The paper reinforces the idea that privacy leaks are often a byproduct of overfitting. By forcing the model to generalize (via Dropout or L2), you naturally hide the specific "fingerprints" of individual training samples.
  • The Model Inversion Paradox: Surprisingly, DP-SGD actually increased the success of model inversion attacks in some scenarios. The noise injected by DP-SGD in the final layers appeared to make the reconstruction of training data easier (lower Mean Squared Error for the attacker) compared to regularized models.
  • L2 as a Precision Shield: L2-regularization was particularly effective at the prediction layer, significantly hindering an attacker's ability to invert the model from its output.

Key Results

  • Utility: DP-SGD caused accuracy drops between 5% and 14.6%. In contrast, Dropout often improved accuracy over the baseline (especially on CIFAR-10) while still providing privacy benefits.
  • MIA Protection: While DP-SGD ($\epsilon=2$) reduced the attacker's success (AUC) by about 13.8%, the combination of L2 and Dropout achieved a comparable 12.93% reduction without the massive performance hit.
  • Inversion Resistance: L2-regularization improved resistance to model inversion by an average of 73.8%, far outperforming DP-SGD, which sometimes made the model more vulnerable to reconstruction.

Limitations and Future Work

The authors acknowledge that while regularization performs well empirically, it lacks the mathematical provability and "privacy budget" $(\epsilon)$ accounting that Differential Privacy provides. DP offers a theoretical guarantee that holds regardless of the attacker's strategy, whereas regularization is an empirical defense that might be bypassed by more sophisticated, future attacks.

Future work could involve testing these trade-offs on much larger, more complex datasets (like ImageNet) or in different domains like Natural Language Processing, where the relationship between memorization and utility might differ.

Summary updated

Abstract

Nowadays, owners and developers of deep learning models must consider stringent privacy-preservation rules of their training data, usually crowd-sourced and retaining sensitive information. The most widely adopted method to enforce privacy guarantees of a deep learning model nowadays relies on optimization techniques enforcing differential privacy. According to the literature, this approach has proven to be a successful defence against several models’ privacy attacks, but its downside is a substantial degradation of the models’ performance. In this work, we compare the effectiveness of the differentially-private stochastic gradient descent (DP-SGD) algorithm against standard optimization practices with regularization techniques. We analyze the resulting models’ utility, training performance, and the effectiveness of membership inference and model inversion attacks against the learned models. Finally, we discuss differential privacy’s flaws and limits and empirically demonstrate the often superior privacy-preserving properties of dropout and l2-regularization.
BibTeX
@inproceedings{lomurno2022utility,
  author = {Eugenio Lomurno and Matteo Matteucci},
  title = {{On the Utility and Protection of Optimization with Differential Privacy and Classic Regularization Techniques}},
  booktitle = {LOD},
  year = {2022},
  url = {https://openreview.net/forum?id=zZRPUvyBRC},
}

Loading related content...

Figure