Cost-effective ensemble models selection using deep reinforcement learning

Birman, Yoni; Hindi, Shaked; Katz, Gilad; Shabtai, Asaf

doi:10.1016/j.inffus.2021.07.011

Cited by 19 publications

(22 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…CETRA builds upon the work of Birman et al [8], which proposed SPIREL, a DRL-based approach for the efficient utilization of ensembles: rather than deploy all detectors at once, the agent dynamically select which additional detectors (if any) to call based on the results of previous ones. While highly effective, SPIREL is hindered by its inability to adapt its policy to achieve specific performance metric goals (e.g., false-positive rate of no more than 1%), or to easily keep these metrics stable in the face of changing data.…”

Section: Proposed Methodsmentioning

confidence: 99%

“…In this section we present the states, actions, and rewards representation of our proposed approach. Our representation closely follows that of [8], as we build upon this base in the following section.…”

Section: Base Methodsmentioning

confidence: 99%

“…Birman at el. [8] proposed a DRL framework for the dynamic sequential allocation of detectors for each sample. Based on the scores assigned by previous detectors, their framework either allocates additional detectors or produces a classification.…”

Section: Reinforcement Learning-based Securitymentioning

confidence: 99%

“…The aforementioned drawbacks make clear the need for a more refined solution, that can weigh the benefit of using each individual learning model in an ensemble against its cost (measured by running time, computing cost, etc.). Such an elegant solution to this problem was recently proposed in [8], where the authors presented SPIREL, a deep reinforcement learning (DRL) framework for the dynamic sequential allocation of detectors for each sample. Based on the scores assigned by previous detectors, SPIREL either allocates additional detectors or produces a classification.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

A Transferable and Automatic Tuning of Deep Reinforcement Learning for Cost Effective Phishing Detection

Orel¹,

Shabtai²,

Katz³

2022

Preprint

Self Cite

View full text Add to dashboard Cite

Many challenging real-world problems require the deployment of ensemblesmultiple complementary learning models-to reach acceptable performance levels. While effective, applying the entire ensemble to every sample is costly and often unnecessary. Deep Reinforcement Learning (DRL) offers a cost-effective alternative, where detectors are dynamically chosen based on the output of their predecessors, with their usefulness weighted against their computational cost.Despite their potential, DRL-based solutions are not widely used in this capacity, partly due to the difficulties in configuring the reward function for each new task, the unpredictable reactions of the DRL agent to changes in the data, and the inability to use common performance metrics (e.g., TPR/FPR) to guide the algorithm's performance. In this study we propose methods for fine-tuning and calibrating DRL-based policies so that they can meet multiple performance goals. Moreover, we present a method for transferring effective security policies from one dataset to another. Finally, we demonstrate that our approach is highly robust against adversarial attacks.

show abstract

Section: Proposed Methodsmentioning

confidence: 99%

Section: Base Methodsmentioning

confidence: 99%

Section: Reinforcement Learning-based Securitymentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

A Transferable and Automatic Tuning of Deep Reinforcement Learning for Cost Effective Phishing Detection

Orel¹,

Shabtai²,

Katz³

2022

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…Multiple ML algorithms are combined to develop an ensemble model. It is proved that the prediction accuracy of an ensemble model for a complex dataset is much higher than a standalone model [5]. This ensemble technique uses a metalearning stage which ensures the highest accuracy [4] [6].…”

Section: Introductionmentioning

confidence: 99%

An Ensemble Model to Minimize Fluctuation Influences on Short-Term Medical Workload Prediction

Mizan,

Taghipour

2022

Scientia Iranica

View full text Add to dashboard Cite

Real-time data are commonly prone to errors due to irregular fluctuations, seasonal biases, and missing values in the data. The erroneous data causes inaccurate forecasting which leads to business loss. Moreover, the concept drift problem is a known problem in time series forecasting that also results in poor forecasting accuracy. The execution time of a machine learning model is also crucial when it is deployed in a real-time environment. This work presents an Adaptive Batched-Ranked Ensemble (ABRE) model that reduces the effect of fluctuation using the timevariant windowing technique. A data aggregation technique is developed and integrated with the offline training phase of the proposed model to tackle the concept drift problem. A meta-model is developed in the online forecasting phase which ensures faster execution for incoming data. The model is implemented for the medical workload prediction after testing and comparing with a few other heterogeneous ensemble models. The comparison results show in terms of the root mean squared error, the proposed model performs at least 65.7% better than the heterogeneous stacked ensemble models on the experimental dataset. Moreover, in comparison to the other standalone models considered in this experiment, the ABRE model reduces the prediction error by approximately 73.6%.

show abstract

A robust malware traffic classifier to combat security breaches in industry 4.0 applications

Ramaiah

Vanmathi

Ravi

2023

Concurrency and Computation

View full text Add to dashboard Cite

SummaryIndustry 4.0 integrates cyber systems, physical devices, and digital networks to automate the industrial process. Many sectors aim to adopt the best practices outlined in Industry 4.0. This indicates well for the future networking of an increasing number of devices. As crucial as intelligent automation is, it is essential that it be protected. The proliferation of Internet‐enabled gadgets could raise vulnerability to a variety of threats, malware among them. Intruders see a synthesis of factors as a chance to carry out their malicious plan. Keeping sensitive data and information protected from malicious software is a high responsibility for all industries. It is critical to have both a trustworthy approach and a large dataset to work with when constructing a malware traffic classifier. Malware's capacity to elude detection by antivirus programs improves with the day. Because this malware has the potential to compromise the entire network, establishing a malware traffic classifier requires a strong approach. As the number of data increases, the classifier has a harder time distinguishing between benign and malicious network entries. As a result, weighing too many factors is a time‐consuming process. To assist with these types of real‐world challenges, we construct an effective hybrid selection component, which is subsequently followed by a neural network classifier in this research. The Malware traffic classifier provided here selects the principal feature using filter and wrapper techniques. The feature columns provided by the feature selection program are used to construct a neural network‐based binary malware classifier. The given malware traffic classification framework was tested using the MTA‐KDD'19 dataset. We set up an experiment in this investigation to examine the way different feature counts perform using a neural‐based classifier. The suggested framework achieves 96.8% accuracy while just considering the bare minimum of five features, which is a substantial increase over alternative methods.

show abstract

Cost-effective ensemble models selection using deep reinforcement learning

Cited by 19 publications

References 22 publications

A Transferable and Automatic Tuning of Deep Reinforcement Learning for Cost Effective Phishing Detection

A Transferable and Automatic Tuning of Deep Reinforcement Learning for Cost Effective Phishing Detection

An Ensemble Model to Minimize Fluctuation Influences on Short-Term Medical Workload Prediction

A robust malware traffic classifier to combat security breaches in industry 4.0 applications

Contact Info

Product

Resources

About