TL;DR
This paper tackles the challenge of data poisoning attacks on machine learning models when the training data is imbalanced. They propose novel methods that effectively overcome the gradient mismatch caused by common data balancing techniques, significantly improving the success rate of these attacks on imbalanced datasets.
What Problem Does This Paper Address?
Machine learning models are increasingly deployed in critical applications, making them targets for adversarial attacks. Targeted Clean-label Data Poisoning Attacks (TCDPA) are particularly concerning because they subtly alter training data with imperceptible, label-consistent modifications to manipulate specific test samples during deployment. A popular TCDPA strategy, gradient-matching, works by aligning the gradient of poisoned samples with the gradient of a target sample. However, this method's effectiveness degrades significantly when applied to imbalanced datasets. This is because standard data balancing techniques, like re-sampling and re-weighting, used to handle class imbalance, introduce a "gradient mismatch" that disrupts the gradient-matching process. This paper aims to bridge this gap by developing TCDPA methods that are robust to these data balancing techniques.
Methods
The core of this paper lies in proposing two modifications to existing gradient-matching TCDPA methods to make them invariant to class balancing techniques.
- Class-Weight-Invariant Witches Brew (CLI-WB): This method reformulates the attack objective to eliminate its dependency on class-specific weights (used in re-weighting) or implicit class probabilities (from re-sampling). By isolating samples from each class and aligning them individually with the target gradient, the attack becomes robust to changes in class importance.
- Class-Weight-Invariant Witches Brew - Projection (CWI-WB-P): This is a computationally optimized version of CLI-WB. Instead of directly aligning gradients, it optimizes the projection of poisoned sample gradients onto the target gradient. This linearizes the objective, reducing the computational cost of the attack optimization without sacrificing effectiveness, and potentially amplifying the poisoning effect.
The authors adapt the Witches Brew algorithm as their baseline and implement their proposed methods by modifying its objective function.
Datasets
The experiments were conducted on the following datasets:
- MNIST and CIFAR10: These standard datasets were artificially made imbalanced using a stratified sub-sampling scheme to create a class imbalance ratio of 0.1 (tail class samples are 10% of head class samples). These are referred to as
si-MNISTandsi-CIFAR10. - MIT-BIH: This is a real-world ECG dataset known for its natural class imbalance. The authors used a coarse classification of heartbeats, with the "Normal" class comprising over 85% of the training samples.
Key Insights and Ideas
- Gradient Mismatch is the Bottleneck: The paper clearly demonstrates that the effectiveness of gradient-matching TCDPA is severely hampered by data balancing techniques due to gradient mismatch. This is a crucial insight for understanding the vulnerabilities of these attacks in realistic scenarios.
- Invariance is Key: The proposed methods achieve robustness by making the attack objective invariant to the class balancing strategies. This is a generalizable principle that could be applied to other adversarial attack scenarios involving imbalanced data.
- Projection for Efficiency: The CWI-WB-P method highlights an interesting trade-off between computational cost and attack effectiveness. Optimizing the projection of gradients offers a practical way to speed up attacks while maintaining or even improving performance.
- AESR as a More Reliable Metric: The authors introduce and advocate for the Attack Effect Success Rate (AESR) metric over the standard Attack Success Rate (ASR) for imbalanced datasets. This is because ASR can be misleading due to inherent low performance on minority classes and the potential bias introduced by label-shift correction methods.
Key Results
- The proposed methods, CWI-WB and CWI-WB-P, consistently outperform the baseline Witches Brew attack on imbalanced datasets.
- Notable improvements were observed: up to 32% (Re-sampling) and 51% (Re-weighting) in terms of Attack Effect Success Rate (AESR) on MNIST and CIFAR10.
- On the MIT-BIH dataset, CWI-WB showed comparable or better attack success rates compared to Witches Brew in scenarios mimicking false negatives and incorrect medical diagnoses.
- The F1-scores of models poisoned using the proposed methods were similar to those of clean models, indicating that the attacks do not significantly degrade the overall model performance on standard test sets.
Limitations and Future Work
- Computational Cost: While CWI-WB-P offers efficiency gains, the original CWI-WB method can still have a higher practical run-time due to the need for multiple differentiation steps per class. The paper notes that a large number of classes can significantly increase this run-time.
- AESR Nuances: While AESR is presented as a better metric, the authors acknowledge that it might not always directly correlate with the number of label flips, suggesting it's more for comparing attack methods than assessing practical impact.
- Future Work: The paper suggests extending the methods to handle more complex scenarios, such as attacks with larger budgets or more sophisticated data balancing techniques. Further exploration into the theoretical underpinnings of why CWI-WB might outperform CWI-WB-P could also be a fruitful direction.
Summary updated