Abstract
Machine Learning models have been shown to be vulnerable to adversarial examples, i.e., the manipulation of data by an attacker to defeat a defender's classifier at test time. We present a novel probabilistic definition of adversarial examples in perfect or limited knowledge settings using prior probability distributions on the defender's classifier. Using the asymptotic properties of logistic regression, we derive a closed-form expression of the intensity of any adversarial perturbation to achieve a given expected misclassification rate. This technique is relevant in a threat model of known model specifications and unknown training data. To our knowledge, this is the first method that allows an attacker to directly choose the probability of attack success. We evaluate our approach on two real-world datasets.
Introduction
In the field of adversarial machine learning, "transferability" describes the phenomenon where an adversarial example designed to fool one model also successfully fools another. Traditionally, researchers have treated transferability as a binary outcome or a stochastic side effect. This paper, however, introduces a method to precisely "dial in" the desired probability of an attack's success ($\alpha$) against a defender's model.
The research addresses a "limited knowledge" threat model where the attacker knows the defender is using logistic regression and knows the specific hyperparameters (such as regularization), but lacks access to the defender's training data. To compensate, the attacker trains a surrogate model on similar data. Because the surrogate's decision boundary will not perfectly match the defender's, a perturbation that barely crosses the surrogate's boundary might fail against the defender. The authors propose a method to calculate the exact additional "push"—or perturbation intensity—needed to ensure a specific success rate.
Methodology: The $\alpha$-Adversarial Example
The core innovation of this work is the definition of the $\alpha$-adversarial example, where $\alpha$ is the user-defined probability of a successful attack. By treating the defender's model parameters as random variables, the authors leverage the fact that logistic regression parameter estimates ($\hat{\theta}$) follow an asymptotic normal distribution as the sample size grows.
The process involves two primary steps:
- Craft the Base Attack: The attacker identifies the optimal $L_2$ perturbation for their own surrogate model.
- Calculate the Intensity ($\lambda$): Using the variance-covariance matrix of the logistic regression parameters, the authors derive a closed-form quadratic equation to find a scaling factor $\lambda$. This factor adjusts the perturbation intensity so that the probability of the perturbed input falling on the incorrect side of the defender's boundary is exactly $\alpha$.
This approach utilizes the Karush–Kuhn–Tucker (KKT) conditions to solve the optimization problem and relies on the quantile function of the normal distribution to map the desired success rate to a physical perturbation distance.
Key Insights and the Regularization Paradox
A counter-intuitive finding of this research is the Regularization Paradox: strong $L_2$ regularization can actually assist the attacker. While regularization is intended to make models more robust, it reduces the variance of the parameter estimates. This makes the defender’s model more "predictable" to the attacker, allowing for smaller, more precise perturbations to achieve the target success rate $\alpha$.
Furthermore, the paper highlights that the specific estimation method (e.g., Iteratively Reweighted Least Squares vs. coordinate descent solvers like Liblinear) significantly alters the variance-covariance structure, which in turn changes the required perturbation intensity. The authors also demonstrate the necessity of Individualized Scaling; because decision boundary uncertainty varies across the feature space, each input requires a custom-calculated $\lambda$ to hit the target $\alpha$.
Experimental Results
The method was evaluated on two binary classification tasks: the UCI Spambase tabular dataset and the Dogs vs. Cats computer vision dataset (downsampled to $64 \times 64$ pixels).
- Precision Control: On the Spambase dataset, the authors demonstrated the ability to successfully calculate the intensity needed to move from a standard attack (typically yielding a ~50% transfer rate) to a high-confidence attack ($\alpha = 0.95$).
- Visual Impact: In the Dogs vs. Cats experiment, the results showed that for images the defender correctly classified, the perturbation intensity must increase as $\alpha$ increases. Conversely, if the defender's model was already likely to misclassify an image, the required additional intensity was zero or even negative.
- Computational Feasibility: Because the derivation provides a closed-form solution, finding the optimal intensity is nearly instantaneous once the variance-covariance matrix is estimated.
Limitations and Future Work
While promising, the method faces a dimensionality bottleneck: it requires inverting a $p \times p$ matrix (where $p$ is the number of features). This is manageable for tabular data but computationally expensive for high-resolution images or deep learning architectures. Additionally, the current derivation is specific to logistic regression. Future work aims to extend this framework to other models with known asymptotic distributions and to "unknown model" scenarios where the attacker does not know the defender's specific model type.
Summary updated