Use the original Enron dataset to create an autoencoder that would identify suspicious from non-suspicious emails. Instead of an autoencoder, you can try with Variational Autoencoders. Compare the performance of both approaches.