Fast Adversarial Label-Flipping Attack on Tabular Data

Xinglong Chang, Gillian Dobbie, Jörg WickerarXiv 2023 · DOI · arXiv

Abstract

FALFA (Fast Adversarial Label-Flipping Attack) is a highly efficient poisoning attack designed to compromise machine learning models by strategically flipping training labels. By transforming a complex, non-linear adversarial objective into a streamlined linear programming problem, FALFA achieves state-of-the-art attack effectiveness on tabular data with a massive reduction in computational overhead. This paper demonstrates that FALFA consistently outperforms existing methods in both speed and degradation of model accuracy across diverse tabular datasets.

Introduction

While adversarial research often focuses on computer vision or natural language processing, tabular data remains the backbone of critical infrastructure, particularly in cybersecurity domains like intrusion and malware detection. In these settings, verifying the "ground truth" of a label often requires deep domain expertise, making it easier for malicious actors to inject "poison" into training sets unnoticed.

The authors address Adversarial Label-Flipping Attacks (ALFAs). Unlike feature-based attacks that require modifying complex data points, ALFAs only require the attacker to change the labels (e.g., changing "malicious" to "benign"). However, existing ALFA methods are often too computationally expensive for high-dimensional tabular data or fail to scale effectively. FALFA bridges this gap by creating an attack that is both devastatingly effective and computationally practical.

Methodology

The core contribution is FALFA, an algorithm derived directly from the adversary’s objective: maximizing the loss on clean test data while minimizing the loss on the poisoned training data to keep the attack "hidden."

Key technical innovations include:

  • Linear Programming Transformation: The authors use a variable transformation technique to convert a highly non-linear objective function—which usually involves absolute operations that are computationally heavy—into a Linear Programming (LP) problem.
  • Complexity Reduction: Unlike previous state-of-the-art methods that effectively double the number of variables in the optimization step, FALFA keeps the variable count lean. This results in an optimization process that is theoretically ~5.6 times faster per iteration.
  • Cross-Entropy Optimization: FALFA is specifically tailored for classifiers that use Cross-Entropy loss, such as Neural Networks, making it a potent threat against modern deep learning architectures.

Experimental Setup

The researchers evaluated FALFA using ten real-world tabular datasets sourced from the UCI Machine Learning Repository, representing a mix of "Easy" (E), "Normal" (N), and "Hard" (H) classification tasks.

Table 1: Characteristics of the 10 real-world tabular datasets used for evaluation, including sample size (n), feature count (m), and baseline accuracies.

Multi-class datasets were converted into binary classification tasks to standardize the attack environment. The datasets vary significantly in dimensionality and baseline performance, providing a robust testbed for the attack's generalizability.

Key Results and Discussion

FALFA demonstrated superior performance across the board. At a 10% poisoning rate, FALFA tied or outperformed existing state-of-the-art attacks (ALFA and PoisSVM) on 7 out of 10 datasets.

Figure 1: Comparison of Accuracy vs. Poisoning Rate across 10 datasets. FALFA (red) consistently shows a sharper decline in test accuracy compared to baseline noise and other attack methods.

The "Camouflage" Effect

A striking finding is that label-flipping can make a difficult dataset appear as an "easily solvable" problem during training. This can trick practitioners into a false sense of security, leading them to lower their defenses just as the model's real-world performance is being sabotaged. While Neural Networks are robust against random label noise, they are highly vulnerable to the strategic adversarial noise generated by FALFA.

Computational Efficiency

The computational gains were significant. For example, on the CMC dataset, FALFA completed the attack in approximately 22 seconds, while the standard ALFA took 406 seconds, and PoisSVM took over 2 hours. FALFA typically converges in just 2 iterations, whereas previous methods often require 5 or more.

Conclusion

FALFA represents a significant advancement in adversarial machine learning for tabular data, offering a fast and effective method for label-flipping attacks. While currently optimized for binary classification and Cross-Entropy loss, its efficiency highlights a critical vulnerability in modern machine learning pipelines. Future work will focus on extending FALFA to multi-class settings and developing robust defense mechanisms, such as label sanitization, to identify these camouflaged adversarial labels before they compromise model integrity.

Summary updated

Abstract

Machine learning models are increasingly used in fields that require high reliability such as cybersecurity. However, these models remain vulnerable to various attacks, among which the adversarial label-flipping attack poses significant threats. In label-flipping attacks, the adversary maliciously flips a portion of training labels to compromise the machine learning model. This paper raises significant concerns as these attacks can camouflage a highly skewed dataset as an easily solvable classification problem, often misleading machine learning practitioners into lower defenses and miscalculations of potential risks. This concern amplifies in tabular data settings, where identifying true labels requires expertise, allowing malicious label-flipping attacks to easily slip under the radar. To demonstrate this risk is inherited in the adversary's objective, we propose FALFA (Fast Adversarial Label-Flipping Attack), a novel efficient attack for crafting adversarial labels. FALFA is based on transforming the adversary's objective and employs linear programming to reduce computational complexity. Using ten real-world tabular datasets, we demonstrate FALFA's superior attack potential, highlighting the need for robust defenses against such threats.
BibTeX
@misc{chang2023fast,
  author = {Xinglong Chang and Gillian Dobbie and Jörg Wicker},
  title = {{Fast Adversarial Label-Flipping Attack on Tabular Data}},
  year = {2023},
  eprint = {2310.10744},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.10744},
}
Figure