Provably effective detection of effective data poisoning attacks

TL;DR

This paper introduces a mathematically rigorous method to detect data poisoning attacks using the Conformal Separability Test, providing both theoretical guarantees and experimental validation of its effectiveness.

Contribution

It defines dataset poisoning attacks precisely and proves they can be effectively detected with a new statistical test, bridging theory and practical detection.

Findings

The Conformal Separability Test reliably detects poisoning in real-world datasets.

Mathematical proof guarantees poisoning detection is always possible if poisoning occurs.

Experimental results confirm the test's effectiveness in practical scenarios.

Abstract

This paper establishes a mathematically precise definition of dataset poisoning attack and proves that the very act of effectively poisoning a dataset ensures that the attack can be effectively detected. On top of a mathematical guarantee that dataset poisoning is identifiable by a new statistical test that we call the Conformal Separability Test, we provide experimental evidence that we can adequately detect poisoning attempts in the real world.