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.
Introduction
Deep learning models are notorious for their tendency to memorize training data rather than simply learning generalizable patterns. This memorization creates significant vulnerabilities, particularly when models are trained on sensitive personal information. Two primary threats have emerged in the privacy landscape: Membership Inference Attacks (MIA), where an attacker determines if a specific individual's data was used in training, and Model Inversion Attacks, where an attacker attempts to reconstruct original training samples by probing the model's internal states.
To combat these threats, Differentially Private Stochastic Gradient Descent (DP-SGD) has become the industry standard. By injecting noise into the gradient computation process, DP-SGD provides a mathematical guarantee of privacy. However, this guarantee often comes at a steep "privacy tax"—a significant reduction in model accuracy and a massive increase in computational overhead. This paper investigates whether classic regularization techniques, such as Dropout and L2-regularization, can provide comparable privacy protections without the associated performance penalties.
Methodology
To evaluate the trade-offs between privacy and utility, we conducted a head-to-head comparison across three standard image benchmarks: MNIST, Fashion-MNIST (F-MNIST), and CIFAR-10. We compared several model configurations:
- Baseline: Standard training without specific privacy or heavy regularization.
- DP-SGD: Training with $(\epsilon, \delta)$-differential privacy at varying privacy budgets ($\epsilon = 2, 4, 8$).
- Regularization: Models utilizing Dropout (20% rate), L2-regularization (weight decay), and a combination of both.
These models were subjected to two classes of attacks. First, we deployed MIA using both threshold-based techniques and shadow-model strategies. Second, we utilized a custom-designed Model Inversion Attack that attempts to reconstruct input data from the model's internal activation maps and final prediction vectors.
The Performance Cost of Privacy
Our findings highlight a dramatic disparity in training efficiency. DP-SGD requires computing gradients for individual "micro-batches" to accurately calibrate noise, a process that resulted in training times 41 to 50 times slower than standard training.
Beyond computational cost, the utility of the models suffered significantly under DP-SGD. We observed accuracy drops ranging from 5% to 14.6% compared to the baseline. In contrast, regularization techniques like Dropout often maintained or even improved model accuracy (particularly on the more complex CIFAR-10 dataset) while still providing a layer of privacy protection.
Empirical Privacy: MIA and Model Inversion
The core of our investigation focused on whether regularization could match the defensive capabilities of DP-SGD. The results suggest that overfitting is a primary driver of privacy leaks; by forcing the model to generalize through Dropout or L2, the specific "fingerprints" of individual training samples are naturally obscured.
Membership Inference Protection
While DP-SGD with a strict privacy budget ($\epsilon=2$) reduced the attacker's success (AUC) by approximately 13.8%, the combination of L2-regularization and Dropout achieved a comparable 12.93% reduction. This suggests that for MIA, the empirical protection offered by regularization is nearly as effective as the formal protection of DP-SGD, but without the massive utility hit.
The Model Inversion Paradox
Perhaps the most surprising finding was the vulnerability of DP-SGD to model inversion. In several scenarios, the noise injected by DP-SGD actually increased the success of reconstruction attacks. The added noise in the final layers appeared to make it easier for the attacker to invert the model, resulting in lower Mean Squared Error for the reconstructed images compared to regularized models.
L2-regularization, specifically when applied to the prediction layer, acted as a "precision shield." It improved resistance to model inversion by an average of 73.8%, far outperforming DP-SGD in this specific defensive category.
Conclusion and Future Work
While DP-SGD remains the only method with a mathematically provable privacy budget ($\epsilon$), our empirical results suggest it may not always be the most practical or effective defense in real-world scenarios. Classic regularization techniques offer a compelling alternative, providing significant resistance to both MIA and model inversion while preserving model utility and training speed.
However, it is important to note that regularization is an empirical defense. Unlike DP, it lacks a theoretical guarantee that holds regardless of the attacker's strategy. Future work should explore these trade-offs on larger, more complex datasets like ImageNet and in the domain of Natural Language Processing to determine if these findings hold as model complexity continues to scale.
Summary updated