Using contrastive learning to improve the performance of steganalysis schemes