Introduction
While adversarial machine learning (AML) has predominantly focused on deep learning architectures—exemplified by the well-known "panda-to-gibbon" image perturbations—classical statistical models have remained relatively unscrutinized. Multivariate Gaussians (MVGs) serve as the foundational "workhorses" of modern science and engineering, underpinning critical applications such as Bayesian optimization, real estate valuation, and Kalman filters for satellite navigation. This paper addresses a significant security gap by exploring how an adversary can subtly corrupt input data to mislead decision-makers relying on these ubiquitous probabilistic models.
Problem Formulation
The core challenge involves disrupting conditional inference. In a typical scenario, a decision-maker observes a subset of variables ($Z$) to predict or infer the state of others ($Y$). An attacker capable of intercepting and modifying $Z$ faces a dual-objective optimization problem:
- Maximize Disruption: The attacker aims to maximize the error in the prediction of $Y$, quantified by the Kullback-Leibler (KL) divergence between the true conditional distribution and the one induced by the corrupted data.
- Maintain Plausibility: To avoid detection, the corrupted data must remain realistic. This is measured by the log-ratio of the marginal density, ensuring the modified observation does not deviate too far from the expected distribution (the mode).
Methodology
The researchers evaluate the attack under two distinct knowledge settings:
White-Box Attacks
In this setting, the attacker possesses perfect knowledge of the model’s parameters, specifically the mean ($\mu$) and the covariance matrix ($\Sigma$). This scenario is framed as a Quadratic Program (QP), which can be solved efficiently using standard optimization solvers like CPLEX.
Grey-Box Attacks
Here, the attacker is uncertain about the exact model parameters and instead utilizes a Bayesian prior, specifically a Normal-Inverse-Wishart distribution. To navigate this uncertainty, the authors employ stochastic programming techniques, including Sample Average Approximation (SAA) and variants of Stochastic Gradient Ascent (SGA) such as Adam and RMSProp.
Empirical Evaluation
The framework was tested across three practical domains to demonstrate the real-world implications of these vulnerabilities:
- Zillow Home Value Index (ZHVI): Using real-world data from Arizona counties, the authors demonstrated how an attacker could disrupt tax assessments or investment decisions. The white-box attack achieved a KL divergence of 255.3, significantly outperforming a random noise baseline of 108.9 while maintaining identical plausibility constraints.
- Lending Club Loan Data: By corrupting predictors such as income and debt-to-income ratios, the attacker manipulated interest rate estimates. While the shift (e.g., from 12.7% to 12.1%) may appear minor, the authors note that such discrepancies result in massive financial shifts when aggregated across thousands of loans over their lifetime.
- Linear Gaussian State-Space Model (LG-SSM): In a simulated 2D object tracking scenario, the attack modified sensor measurements to deceive a tracker. The results revealed a strategic trade-off: the adversary could either make an object appear on the correct path but at the wrong speed, or at the correct speed but on the wrong path.
Key Insights and Attack Characteristics
The "Interior" Attack
Unlike many deep learning attacks that push data toward the extreme edges of a feasible region (e.g., Fast Gradient Sign Method), these MVG attacks often find optimal points within the interior of the distribution. Moving too far toward a boundary drastically lowers plausibility, making the attack easier to detect.
Grey-Box Aggression
Counter-intuitively, grey-box attacks can sometimes be more aggressive than white-box attacks. The attacker’s "average" belief about the model parameters can lead to a solution that is more disruptive to the true model than a solution derived from exact parameter knowledge.
Obfuscation as Defense
The research suggests that keeping specific model parameters (like the $\Sigma$ matrix) secret can serve as a defense mechanism. Forcing an attacker into the grey-box regime increases the complexity of the attack, though it does not eliminate the threat. Furthermore, the computational efficiency of these attacks is high; even complex grey-box attacks using 10,000 samples were solved in under 10 seconds, making them viable for real-time disruption.
Conclusion and Future Work
This study highlights the inherent vulnerability of classical statistical models to sophisticated adversarial corruption. While the current framework relies on the mathematical properties of Gaussians (such as the Schur complement), future work may extend these methods to non-Gaussian distributions using variational approximations. Additionally, while this paper focuses on indiscriminate disruption, future research could explore targeted attacks designed to steer model outcomes toward specific, beneficial results for the adversary. The ultimate goal remains the development of adversarially robust classical models capable of detecting or mitigating these subtle data corruptions.
Summary updated