The invention discloses a data security deduplication method based on an auto-encoder, relates to the field of information security and artificial intelligence, solves the problem of low efficiency ofan existing data deduplication method based on random message lock encryption, introduces abstract tags in efficiency, quickly screens out a very small subset from a tag library by means of the similarity of the tags, and executes bilinear mapping calculation on the subset, so that the frequency of bilinear mapping calculation is greatly reduced, and the label comparison efficiency is improved. According to the method, a self-encoding technology commonly used in image processing is introduced, the deduplication efficiency is improved by greatly reducing the number of times of label comparison, and the deduplication efficiency is improved by nearly 10 times compared with a data deduplication method based on random message lock encryption. According to the method, the non-monotonic functionis introduced, so that similar labels can be possibly generated even if data with large difference exists, namely, the similar labels can be generated by the similar data but cannot be established inturn, and the difficulty of deducing the data by an attacker according to the labels is further improved.