Home → Magazine Archive → February 2023 (Vol. 66, No. 2) → Proving Data-Poisoning Robustness in Decision Trees → Abstract

Proving Data-Poisoning Robustness in Decision Trees

By Samuel Drews, Aws Albarghouthi, Loris D'Antoni

Communications of the ACM, Vol. 66 No. 2, Pages 105-113

[article image]

Machine learning models are brittle, and small changes in the training data can result in different predictions. We study the problem of proving that a prediction is robust to data poisoning, where an attacker can inject a number of malicious elements into the training set to influence the learned model. We target decision tree models, a popular and simple class of machine learning models that underlies many complex learning techniques. We present a sound verification technique based on abstract interpretation and implement it in a tool called Antidote. Antidote abstractly trains decision trees for an intractably large space of possible poisoned datasets. Due to the soundness of our abstraction, Antidote can produce proofs that, for a given input, the corresponding prediction would not have changed had the training set been tampered with or not. We demonstrate the effectiveness of Antidote on a number of popular datasets.

Back to Top

1. Introduction

Artificial intelligence, in the form of machine learning (ML), is rapidly transforming the world as we know it. Today, ML is responsible for an ever-growing spectrum of sensitive decisions—from loan decisions, to diagnosing diseases, to autonomous driving. Many recent works have shown how ML models are brittle,2,5,19,20,22 and with ML spreading across many industries, the issue of robustness in ML models has taken center stage. The research field that deals with studying robustness of ML models is referred to as adversarial machine learning. In this field, researchers have proposed many definitions that try to capture robustness to different adversaries. The majority of these works have focused on verifying or improving the model's robustness to test-time attacks,8,9,21 where an adversary can craft small perturbations to input examples that fool the ML model into changing its prediction, for example, making the model think a picture of a cat is that of a zebra.4


No entries found