Optimization with neural networks trained by evolutionary algorithms

Velazco, Marta; Filho, Christiano Lyra

doi:10.1109/ijcnn.2002.1007742

Cited by 4 publications

(2 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The layers in between are termed as hidden layers. The number of hidden layers varies depending on the design [50]. The structure of the neural network is shown in Fig.…”

Section: Artificial Neural Networkmentioning

confidence: 99%

Analysis of Swarm Intelligence Based ANN Algorithms for Attacking PUFs

2021

View full text Add to dashboard Cite

Physical Unclonable Functions (PUFs) are used for authentication and generation of secure cryptographic keys. However, recent research work has shown that PUFs, in general, are vulnerable to machine learning modeling attacks. From a subset of Challenge-Response Pairs (CRPs), the remaining CRPs can be effectively predicted using different machine learning algorithms. In this work, Artificial Neural Networks (ANNs) using swarm intelligence-based modeling attacks are used against different silicon-based PUFs to test their resiliency to these attacks. Amongst the swarm intelligence algorithms, the Gravitational Search Algorithm (GSA), Cuckoo Search Algorithm (CS), Particle Swarm Optimizer (PSO) and the Grey Wolf Optimizer (GWO) are used. The attacks are extensively performed on six different types of PUFs; namely, Configurable Ring Oscillator, Inverter Ring Oscillator, XOR-Inverter Ring Oscillator, Arbiter, Modified XOR-Inverter Ring Oscillator, and Hybrid Delay Based PUF. From the results, it can be concluded that the first four PUFs under study are vulnerable to ANN swarm intelligence-based models, and their responses can be predicted with an average accuracy of 71.1% to 88.3 % for the different models. However, for the Hybrid Delay Based PUF and the Modified XOR-Inverter Ring Oscillator PUF, which are especially designed to thwart machine learning attacks, the prediction accuracy is much lower and in the range of 9.8 % to 14.5 %.

show abstract

“…The layers in between are termed as hidden layers. The number of hidden layers varies depending on the design [50]. The structure of the neural network is shown in Fig.…”

Section: Artificial Neural Networkmentioning

confidence: 99%

Analysis of Swarm Intelligence Based ANN Algorithms for Attacking PUFs

2021

View full text Add to dashboard Cite

show abstract

“…In the forward phase, the outputs for given set of inputs are predicted and in the backward phase gradient of the error is propagated backward and adjusting the weights and the biases in the process. The initial weights and biases are randomly chosen [26] [27]. However, new evolutionary algorithms are emerging that train the weights and biases using evolutionary approach.…”

Section: Artificial Neural Networkmentioning

confidence: 99%

Immuno-Computing-based Neural Learning for Data Classification

Bataineh¹,

Kaur²

2019

IJACSA

View full text Add to dashboard Cite

The paper proposes two new algorithms based on the artificial immune system of the human body called Clonal Selection Algorithm (CSA) and the modified version of Clonal Selection Algorithm (MCSA), and used them to train the neural network. Conventional Artificial Neural Network training algorithm such as backpropagation has the disadvantage that it can get trapped into the local optima. Consequently, the neural network is usually incapable of obtaining the best solution to the given problem. In the proposed new CSA algorithm, the initial random weights chosen for the neural networks are considered as a foreign body called an antigen. As the human body creates several antibodies in response to fight the antigen, similarly, in CSA algorithm antibodies are created to fight the antigen. Each antibody is evaluated based on its affinity and clones are generated for each antibody. The number of clones depends on the algorithm, in CSA, the number of clones is fixed and in MCSA, number of clones is directly proportional to the affinity of the antibody. Mutation is performed on clones to improve the affinity. The best antibody emerged becomes the antigen for the next round and the process is repeated for several iterations until the best antibody that satisfies the chosen criterion is found. The best antibody is problem specific. For neural network training for data classification, the best antibody represents the set of weights and biases that gives the least error. The efficiency of the algorithm was analyzed using Iris dataset. The prediction accuracy of the algorithms were compared with other natureinspired algorithms, such as Ant Colony Optimization (ACO), Particle Swarm Optimization (PSO) and standard backpropagation. The performance of MCSA was ahead of other algorithms with an accuracy of 99.33%.

show abstract