Phishing is a cyber-attack that intends to trick individuals into providing an attacker with sensitive information, such as login authorizations or business information. One tactic that attackers use is creating fake websites that mimic legitimate ones to fool victims into entering their information. A learning-based model can be used to analyze patterns in website content, structure, behavior, email text, sender, and links. These models can help identify phishing attempts and protect individuals and organizations from falling victim to these scams. In this research, we study and experiment with the deep learning-based algorithm in classifying phishing webpages from legitimate webpages that can be generalized across multiple domains. We used the open-source benchmark dataset, "Phishing Dataset for Machine Learning," available on Kaggle. We propose a hybrid and a multi-level ensemble approach for phishing website detection. Several one-class classifiers are trained on the first level, and the Variational Autoencoder (VAE) is used to reduce the size of feature vectors. Each one class classifiers have its own strength and limitations, and thus an adaptive weightage approach is applied. At the second level, a multilayer neural network is trained for classification. Further, the training time is reduced due to feature reduction in the latent space. Our experimental results classify phishing websites from legitimate websites with 98% accuracy.