Steering Bayesian Inference and Decisions through Data Poisoning
Authors: Roi Naveiro, William N. Caballero, Juan Maroñas
Organizations: CUNEF Universidad, Madrid, Spain · Air Force Institute of Technology, OH, USA · University of Arkansas, AR, USA · Machine Learning Group, Universidad Autónoma de Madrid, Spain
Adversarial machine learning has revealed that statistical models can be manipulated through poisoned data, and Bayesian inference is no exception. Earlier work has shown that subtle poisoning attacks can drive the posterior toward a distribution of the attacker's choosing. Such posterior attraction attacks require a fully specified target posterior; however, attackers typically care about only a few posterior quantities or, more consequentially, about a downstream decision. Herein, we study how an attacker can alter Bayesian inference and decisions by deleting or replicating a few observations, chosen to minimize the maximum mean discrepancy between the resulting posterior and a target distribution. A suitable kernel lets the attacker steer selected posterior moments without specifying the rest of the target. Because Bayes actions depend on the posterior only through expected utilities, which are themselves posterior moments, the same approach can target a decision directly. We construct the attacks with a gradient-based heuristic that requires only posterior samples, and analyze its convergence. Experiments with regression and hierarchical models compare moment-based and full-posterior attacks, and illustrate how deleting a handful of observations can flip a Bayesian decision.
Figures & tables
g^(θ)←N1∑θ′∈Sestk(θ,θ′)−M1∑θ′∈SAk(θ,θ′)
Algorithm 1 Unbiased Gradient Estimator for MPA
Figure 1: Linear-regression attack as a function of the perturbation budget B , by heuristic and kernel.
Figure 2: Posterior movement at budget B=80 for Adam-R2: clean, attacked, and target posteriors of (β1,β2) under the moment-matching (left) and RBF (right) kernels. For each kernel, the run with the smallest MMD 2 is shown.
Figure 3: Sensitivity of the targeted posterior mean, under the Adam-R2 attack, to the population covariate correlation ρ of the data-generating process, with unit covariate variances and one simulated dataset per ρ . The dashed line marks the target value 0 .
Figure 4: Achieved posterior mean of βFFREQ as a function of the deletion budget B . Points are averages over five runs; error bars show ±2 standard errors.
Figure 5: Posterior comparison across regression coefficients for the selected budget- 20 attack (Adam-R2).
Figure 6: Posterior distribution of the Lake County utility gap g1(θ) under the clean data and after the budget- 7 attack. Positive values favor remediation ( d1 ); negative values favor no action ( d2) .
Figure 7: Unbiasedness of the stochastic MMD 2 gradient estimator at w=1 , over nrep=20000 replications. Left: Monte Carlo average of the estimator against the exact gradient, for each coordinate ( ±2 standard errors). Middle: standardized discrepancy zi ; the shaded band is ∣zi∣<2 . Right: running mean of the estimator at the coordinate with the largest absolute gradient; the dashed line is the exact value.
Figure 8: Variance attack on the linear-regression posterior as a function of the perturbation budget B , by heuristic and kernel. The dashed line marks the adversarial target σA2 ; dash-dotted lines mark the clean posterior values.
Figure 9: Posterior movement at budget B=80 for Adam-R2: clean, induced, and target posteriors of (β1,β2) under the moment-matching (left) and RBF (right) kernels. In the moment-matching panel, the dotted line marks the clean mean βˉ1 , which the attack must preserve, and the shaded band spans βˉ1±σA , one target standard deviation on each side of it ( σA2=0.5 ); no ellipse is drawn because the moment kernel does not define a full adversarial distribution. For each kernel, the run with the smallest MMD 2 is shown.
Figure 10: Variance attack under increasing predictor correlation. Each panel fixes the population correlation ρ of the design and shows the attained Vθ∼πw[β1] against the budget, for both kernels. The adversarial target, marked by the dashed lines, is σA2=0.5 in every panel.
Figure 11: Footprint of the selected budget- 20 Adam-R2 deletion attack. Marker size is proportional to zinc concentration; circled markers are removed observations.
Figure 12: Support-score diagnostics for the selected budget- 20 attack.
Figure 13: Mechanism of the restricted radon attack.
Test-time guidance is a widely used mechanism for steering pretrained diffusion models toward outcomes specified by a reward function. Existing approaches, however, focus on maximizing reward rather than sampling from the true Bayesian posterior, leading to miscalibrated inference. In this work, we show that common test-time guidance methods do not recover the correct posterior distribution and identify the structural approximations responsible for this failure. We then propose consistent alternative estimators that enable calibrated sampling from the Bayesian posterior. We significantly outperform previous methods on a set of Bayesian inference tasks, and set a new state-of-the-art PSNR in black hole image reconstruction.
Daniel Geyfman, Felix Draxler, Jan Groeneveld +3
Department of Computer Science, University of California, Irvine, CA, USA · Seoul National University, Seoul, South Korea · Chan-Zuckerberg Initiative, Redwood City, CA, USA
Theoretical and applied research into privacy encompasses an incredibly broad swathe of differing approaches, emphases and aims. This work introduces a novel quantitative notion of privacy that is both contextual and specific. Building on and extending ideas from statistical disclosure control and differential privacy, our aim is to model the implications of a disclosure decision in an adversarial setting. Our definition relies on concepts inherent to standard Bayesian decision theory, while departing from them in several important respects. In particular, (i) inference about the data itself becomes meaningful and (ii) the party controlling the release of sensitive information should make disclosure decisions from the prior viewpoint, rather than conditional on the data, which is a feature shared with Bayesian design. Illuminating toy examples are exploited towards highlighting the specificities of the method.
Cameron Bell, Timothy Johnston, Antoine Luciano +1
CEREMADE, Université Paris Dauphine–PSL, France · Department of Statistics, University of Warwick, UK
Scientific experiments are often designed to maximize information gain, yet in many applications the primary objective is to support reliable downstream decision-making. Existing decision-aware experimental design and active learning methods typically assume well-specified outcome models and implicitly rely on the stability of the optimal decision under real-world perturbations. In practice, however, experimental outcomes are frequently influenced by hidden or weakly modeled effects, which can substantially alter decision optimality and lead to misleading conclusions. We study sequential adversarially robust decision-aware experimental design, where data acquisition has to take into account information gain against plausible worst-case unexpected effects, modeled here as variation in adversarial variables. Building on Bayesian decision theory, we formalize an adversarially robust optimal decision under this setting and derive a principled Bayesian experimental design criterion. The criterion explicitly targets decision stability rather than nominal optimality. Experiments on synthetic and real-world scientific datasets show that conventional decision-aware design can converge rapidly to high confidence yet fragile decisions, while our robustness-aware approach yields decisions that are significantly more stable and reliable under adversarial variation.
Department of Computer Science, The University of Manchester, UK · ELLIS Institute Finland · Department of Computer Science, Aalto University, Espoo, Finland