Detection and Mitigation of Label-Flipping Attacks in Federated Learning Systems with KPCA and K-Means

Li, Dongcheng; Wong, W. Eric; Wang, Wei; Yao, Yao; Chau, Matthew

doi:10.1109/dsa52907.2021.00081

Cited by 60 publications

(31 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Here, notation D denotes the total number of aggregations and d l represents the number of times the l-th participant is judged to be a benign participant historically. Importantly, some excellent references [27][28][29] use the clustering method to defend the poisoning attacks. However, the clustering method is only one evidence of our proposed D-S evidence theorybased defense strategy in our paper.…”

Section: Evidence Definitionsmentioning

confidence: 99%

Efficient poisoning attacks and defenses for unlabeled data in DDoS prediction of intelligent transportation systems

Jiang

2022

Security and Safety

View full text Add to dashboard Cite

Nowadays, large numbers of smart sensors (e.g., road-side cameras) which communicate with nearby base stations could launch distributed denial of services (DDoS) attack storms in intelligent transportation systems. DDoS attacks disable the services provided by base stations. Thus in this paper, considering the uneven communication traffic flows and privacy preserving, we give a hidden Markov model-based prediction model by utilizing the multi-step characteristic of DDoS with a federated learning framework to predict whether DDoS attacks will happen on base stations in the future. However, in the federated learning, we need to consider the problem of poisoning attacks due to malicious participants. The poisoning attacks will lead to the intelligent transportation systems paralysis without security protection. Traditional poisoning attacks mainly apply to the classification model with labeled data. In this paper, we propose a reinforcement learning-based poisoning method specifically for poisoning the prediction model with unlabeled data. Besides, previous related defense strategies rely on validation datasets with labeled data in the server. However, it is unrealistic since the local training datasets are not uploaded to the server due to privacy preserving, and our datasets are also unlabeled. Furthermore, we give a validation dataset-free defense strategy based on Dempster–Shafer (D–S) evidence theory avoiding anomaly aggregation to obtain a robust global model for precise DDoS prediction. In our experiments, we simulate 3000 points in combination with DARPA2000 dataset to carry out evaluations. The results indicate that our poisoning method can successfully poison the global prediction model with unlabeled data in a short time. Meanwhile, we compare our proposed defense algorithm with three popularly used defense algorithms. The results show that our defense method has a high accuracy rate of excluding poisoners and can obtain a high attack prediction probability.

show abstract

Section: Evidence Definitionsmentioning

confidence: 99%

Efficient poisoning attacks and defenses for unlabeled data in DDoS prediction of intelligent transportation systems

Jiang

2022

Security and Safety

View full text Add to dashboard Cite

show abstract

“…For example, Wikipedia's introduction to a country, such as Japan, includes the following attributes: 'Climate', 'Biodiversity', 'Environment', 'Agriculture and Fishery', 'Industry', 'Science and Technology', 'Art and Architecture', 'Etiquette', 'Cuisine' and so on. However, Wiki has already grouped these attributes: 'Climate', 'Biodiversity', 'Environment'; these factors are equivalent to𝑢 1 (1) , 𝑢 2 (1) , and, 𝑢 3 (1) in Figure 1, which describe the attribute 'Geography'; thus, we consider 𝑢 1 (1) , 𝑢 2 (1) , and 𝑢 3 (1) to be three child factors of 𝑈 1 , i.e., the parent factor "Geography"; in the same way, 'Agriculture and Fishery', 'Industry', 'Science and Technology' , these factors are equivalent to 𝑢 1 (2) , 𝑢 2 (2) , and, 𝑢 3 (2) in Figure 1, which describe the attribute 'Economy'. So, we consider 𝑢 1 (2) , 𝑢 2 (2) , and 𝑢 3 (2) to be three child factors of 𝑈 2 , i.e., the parent factor "Economy"; finally, 'Art and Architecture', 'Etiquette', 'Cuisine', these factors are equivalent to 𝑢 1 (3) , 𝑢 2 (3) , and, 𝑢 3 (3) in Figure 1, which describe the attribute 'Culture'; therefore, we consider 𝑢 1 (3) , 𝑢 2 (3) , and 𝑢 3 (3) to be three child factors of 𝑈 3 , i.e., the parent factor "Culture", as shown in the Figure 2.…”

Section: Data Preprocessingmentioning

confidence: 99%

“…When we deal with a data set with a large number of attributes, if the values of these attributes are not specific or clearly defined, the amount of calculation in data training will be large and the classification effect will be poor. At the same time, even though the values of each attribute are accurately defined, due to the large number of attributes, it is inevitable that the weight assigned to some attributes will be very small, and the loss of information will be inevitable [1][2]. In that case, the accuracy of the results would be negatively affected when the random forest method is directly used to train such data sets [3].…”

Section: Introductionmentioning

confidence: 99%

An Algorithm Combining Random Forest Classification and Fuzzy Comprehensive Evaluation

Zhu¹,

Zhu²,

Huang³

et al. 2022

Journal of Internet Technology

View full text Add to dashboard Cite

Random forest algorithm is a common classification method. However, if the weights of many attributes in a data set are not same or close to each other, the direct use of this algorithm for data training will lead to the neglect of the interrelationships between these attributes, and it is difficult to reflect the differences brought by different weights of different attributes. Worse, if the number of attributes in the data set is relatively large, many attributes will be given very little weight when normalization is satisfied, which will also lead to information loss. All of these will have a negative impact on the final result. To solve these problems, this paper proposes an algorithm combining random forest classification and fuzzy comprehensive evaluation, which not only take into account the correlation between attributes in data training, but also retain the information in the original data set to the maximun. At the same time, this algorithm significantly improves the accuracy of random forest training results.

show abstract

“…Zhang et al [28] developed an artificial neural network to detect abnormal temperatures of WSNs (Wireless Sensors Networks) in intelligent buildings. Li et al [29] proposed an improved defense strategy that emphasizes employing KPCA and K-means clustering to defend against data-poisoning attacks in federated-learning systems. Bettencourt et al [30] identified fault nodes through the spacetime structure of sensors and neighbor measurements.…”

Section: Anomaly Detectionmentioning

confidence: 99%

Anomaly Detection of Dam Monitoring Data based on Improved Spectral Clustering

Ji¹,

Ji²,

Zhang³

et al. 2022

Journal of Internet Technology

View full text Add to dashboard Cite

In response to the abnormal data mining in dam safety monitoring, and based on the traditional spectral clustering, this paper presents an anomaly detection method based on improved spectral clustering. This method applies a distance and density adaptive similarity measure. The natural eigenvalue is introduced to adaptively select the neighbors of data points, and the similarity is redefined to be combined with the natural k-nearest neighbor. Furthermore, the shared neighbor is introduced to adjust the similarity between the monitoring data samples according to the regional density. Moreover, considering the distribution of dam monitoring data, the initialization of clustering centers is optimized according to both the density and distance feature. This method can prevent the algorithm from local optimum, better adapt to the density of non-convex dataset, reduce the number of iterations, and enhance the efficiencies of clustering and anomaly detection. Taking the dam slab monitoring data as the research object, experimental datasets are formed. Experiments on these datasets further verify that the method of this paper can effectively adapt to discrete distribution datasets and is superior to the classical spectral clustering method in both clustering and anomaly detection.

show abstract

Detection and Mitigation of Label-Flipping Attacks in Federated Learning Systems with KPCA and K-Means

Cited by 60 publications

References 16 publications

Efficient poisoning attacks and defenses for unlabeled data in DDoS prediction of intelligent transportation systems

Efficient poisoning attacks and defenses for unlabeled data in DDoS prediction of intelligent transportation systems

An Algorithm Combining Random Forest Classification and Fuzzy Comprehensive Evaluation

Anomaly Detection of Dam Monitoring Data based on Improved Spectral Clustering

Contact Info

Product

Resources

About