Active Learning Under Malicious Mislabeling and Poisoning Attacks

Lin, Jing; Luley, Ryan; Xiong, Ke

doi:10.48550/arxiv.2101.00157

Cited by 2 publications

(2 citation statements)

References 24 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In our further study, we may apply our proposed consistency check to local model weights for defending against such data poisoning attacks. More precisely, we may follow the consistency check framework proposed in [19].…”

Section: Discussionmentioning

confidence: 99%

Mahalanobis distance-based robust approaches against false data injection attacks on dynamic power state estimation

Lin,

Xiong

2021

Preprint

Self Cite

View full text Add to dashboard Cite

State estimation plays a crucial role in the daily operation of power systems. Although many researchers have studied false data injection (FDI) attacks in power state estimation, current state estimation approaches are still highly vulnerable to FDI attacks. Currently, most existing studies on FDI attacks focus on static state estimation (SSE), where power system states are not changed with time, and one of them includes the discovery of three efficient FDI attacks that can introduce arbitrary large errors into certain state variables without being detected by existing bad measurement detection algorithms. In reality, however, power states are varied with time in real-world power systems. In this paper, we investigate the problem of the above three FDI attacks against dynamic power state estimation (DSE), where power system states are dynamically changed with time. Although the three attacks were discovered in SSE several years ago, none of them has been well addressed in static power state systems. In this research, we propose two robust defense approaches against the above three efficient FDI attacks on DSE. Compared to existing approaches, our proposed approaches have three major differences and significant strengths: (1) they defend against the three FDI attacks on dynamic power state estimation rather than static power state estimation, (2) they give a robust estimator that can accurately extract a subset of attack-free sensors for power state estimation, and (3) they adopt the little-known Mahalanobis distance in the consistency check of power sensor measurements, which is different from the Euclidean distance used in all the existing studies on power state estimation. We mathematically prove that the Mahalanobis distance is not only useful but also much better than the Euclidean distance in the consistency check of power sensor measurements. Our time complexity analysis shows that the proposed two robust defense approaches are efficient. Moreover, in order to demonstrate the effectiveness of the proposed approaches, we compare them with the three wellknown approaches: the least square approach, the Imhotep-SMT approach, and the MEE-UKF approach. Experiments show that the proposed approaches further reduce the estimation error by two orders of magnitude and four orders of magnitude compared to the Imhotep-SMT approach and the least square approach, respectively. Whereas our approach has a similar root mean square error as the Imhotep-SMT approach has, our approach provides a more stable result.

show abstract

Section: Discussionmentioning

confidence: 99%

Mahalanobis distance-based robust approaches against false data injection attacks on dynamic power state estimation

Lin,

Xiong

2021

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…For recent treatments of this topic and further references, see [36,31]. A few recent works have considered data poisoning in the active learning setting [26,39], with defenses focusing on modifying the setting rather than the algorithm. Further, some existing works in active learning regime [24,14] consider the presence of label noise, out-of-distribution examples, and redundancy in the dataset.…”

Section: Related Workmentioning

confidence: 99%

Understanding and Improving Neural Active Learning on Heteroskedastic Distributions

Khosla,

Kin Whye,

Ash

et al. 2023

Frontiers in Artificial Intelligence and Applications

View full text Add to dashboard Cite

Models that can actively seek out the best quality training data hold the promise of more accurate, adaptable, and efficient machine learning. Active learning techniques often tend to prefer examples that are the most difficult to classify. While this works well on homogeneous datasets, we find that it can lead to catastrophic failures when performed on multiple distributions with different degrees of label noise or heteroskedasticity. These active learning algorithms strongly prefer to draw from the distribution with more noise, even if their examples have no informative structure (such as solid color images with random labels). To this end, we demonstrate the catastrophic failure of these active learning algorithms on heteroskedastic distributions and propose a fine-tuning-based approach to mitigate these failures. Further, we propose a new algorithm that incorporates a model difference scoring function for each data point to filter out the noisy examples and sample clean examples that maximize accuracy, outperforming the existing active learning techniques on the heteroskedastic datasets. We hope these observations and techniques are immediately helpful to practitioners and can help to challenge common assumptions in the design of active learning algorithms. Our code is available at https://github.com/savya08/Active-Learning-on-Heteroskedastic-Distributions.

show abstract

Active Learning Under Malicious Mislabeling and Poisoning Attacks

Cited by 2 publications

References 24 publications

Mahalanobis distance-based robust approaches against false data injection attacks on dynamic power state estimation

Mahalanobis distance-based robust approaches against false data injection attacks on dynamic power state estimation

Understanding and Improving Neural Active Learning on Heteroskedastic Distributions

Contact Info

Product

Resources

About