[Paper Review] Poisoning Attacks with Generative Adversarial Nets

Luis Muñoz-González, Bjarne Pfitzner+3 more|arXiv (Cornell University)|Jun 18, 2019

Adversarial Robustness in Machine LearningComputer Science28 references39 citations

TL;DR

The paper introduces pGAN, a GAN-based framework with a generator, discriminator, and target classifier to craft poisoning points that degrade a classifier while remaining detectable.

ABSTRACT

Machine learning algorithms are vulnerable to poisoning attacks: An adversary can inject malicious points in the training dataset to influence the learning process and degrade the algorithm's performance. Optimal poisoning attacks have already been proposed to evaluate worst-case scenarios, modelling attacks as a bi-level optimization problem. Solving these problems is computationally demanding and has limited applicability for some models such as deep networks. In this paper we introduce a novel generative model to craft systematic poisoning attacks against machine learning classifiers generating adversarial training examples, i.e. samples that look like genuine data points but that degrade the classifier's accuracy when used for training. We propose a Generative Adversarial Net with three components: generator, discriminator, and the target classifier. This approach allows us to model naturally the detectability constrains that can be expected in realistic attacks and to identify the regions of the underlying data distribution that can be more vulnerable to data poisoning. Our experimental evaluation shows the effectiveness of our attack to compromise machine learning classifiers, including deep networks.

Motivation & Objective

• Motivate data poisoning as a security threat in machine learning and assess realistic attack constraints.
• Propose a scalable poisoning strategy that works with deep networks by using Generative Adversarial Nets.
• Incorporate detectability controls to model realistic attacker constraints and study trade-offs between attack effectiveness and stealth.

Proposed method

• Introduce pGAN with three components: generator, discriminator, and target classifier.
• Formulate a minimax game where the generator maximizes a convex combination of attacks on the classifier and evasion of the discriminator.
• Use a parameter alpha to weight detectability versus effectiveness, and a poisoning fraction lambda to control injected points.
• Train via coordinated gradient-based updates in a conditional GAN-like setup, conditioned on poisoning class labels.
• Allow surrogate models for black-box scenarios and incorporate standard GAN stabilization techniques (dropout, batch norm, label smoothing).
• Provide practical guidelines on training dynamics, including the role of lambda and the saddle-point solution for the minimax objective.

Experimental results

Research questions

• RQ1 Can a GAN-based framework generate poisoning examples that degrade classifier performance while remaining close to genuine data?
• RQ2 How do detectability constraints (via alpha) affect poisoning effectiveness and stealth?
• RQ3 What is the impact of the poisoning fraction lambda on attack success across datasets and models?
• RQ4 Can pGAN produce targeted, error-specific attacks without excessively increasing detectability?
• RQ5 How does pGAN compare to traditional poisoning methods under detectability constraints?

Key findings

• pGAN can degrade classifier accuracy on MNIST and Fashion-MNIST as poisoning points are injected, with greater effect at lower alpha values.
• The attack remains harder to detect when alpha is high, illustrating a trade-off between stealth and impact.
• Increasing the fraction of poisoning points generally increases attack effectiveness, but larger datasets reduce the relative impact of poisoning.
• pGAN can perform targeted, error-specific attacks (e.g., increasing misclassification of digit 3 as 5) with small poisoning fractions.
• Compared to label-flipping strategies with detectability constraints, pGAN achieves higher attack effectiveness and different error profiles (more targeted, fewer false positives).
• Attack effectiveness diminishes as the training dataset size grows, but targeted attacks remain feasible on larger models.