The　compression　of　the　ciphertext　of　a　cryptosystem　is　desirable　considering　the　dramatic　increase　in　secure　data　transfer　via　Internet.　In　this　paper,　we　propose　a　simple　and　universal　scheme　to　compress　and　decompress　the　ciphertext　of　an　optical　cryptosystem　by　the　aid　of　deep　learning　(DL).　For　compression,　the　ciphertext　is　first　resized　to　a　relatively　small　dimension　by　bilinear　interpolation　and　thereafter　condensed　by　the　JPEG2000　standard.　For　decompression,　a　well-trained　deep　neural　network　(DNN)　can　be　employed　to　perfectly　recover　the　original　ciphertext,　in　spite　of　the　severe　information　loss　suffered　by　the　compressed　file.　In　contrast　with　JPEG2000　and　JPEG,　our　proposal　can　achieve　a　far　smaller　size　of　the　compressed　file　(SCF)　while　offering　comparable　decompression　quality.　In　addition,　the　SCF　can　be　further　reduced　by　compromising　the　quality　of　the　recovered　plaintext.　It　is　also　shown　that　the　compression　procedure　can　provide　an　additional　security　level,　and　this　may　offer　new　insight　into　the　compressive　encryption　in　optical　cryptosystems.　Both　simulation　and　experimental　results　are　presented　to　demonstrate　the　proposal.