Resting-state　functional　connectivity　(FC)　changes　dynamically　and　major　depressive　disorder　(MDD)　has　abnormality　in　functional　connectivity　networks　(FCNs),　but　few　existing　resting-state　fMRI　study　on　MDD　utilizes　the　dynamics,　especially　for　identifying　depressive　individuals　from　healthy　controls.　In　this　paper,　we　propose　a　methodological　procedure　for　differential　diagnosis　of　depression,　called　HN3D.　which　is　based　on　high-order　functional　connectivity　networks　(HFCN).　Firstly,　HN3D　extracts　time　series　by　independent　component　analysis,　and　partitions　them　into　overlapped　short　series　by　sliding　time　window.　Secondly,　it　constructs　a　FCN　for　each　time　window　and　concatenates　correlation　matrices　of　all　FCNs　to　generate　correlation　time　series.　Then,　correlation　time　series　are　grouped　into　different　clusters　and　high-order　correlations　for　HFCN　is　calculated　based　on　their　means.　Finally,　graph　based　features　of　HFCNs　are　extracted　and　selected　for　a　linear　discriminative　classifier.　Tested　on　21　healthy　controls　and　20　MDD　patients,　HN3D　achieved　its　best　100%　classification　accuracy,　which　is　much　higher　than　results　based　on　stationary　FCNs.　In　addition,　most　discriminative　components　of　HN3D　locate　in　default　mode　network　and　visual　network,　which　are　consistent　with　existing　stationary-based　results　on　depression.　Though　HN3D　needs　to　be　studied　further,　it　is　helpful　for　the　differential　diagnosis　of　depression　and　might　have　potentiality　in　identifying　relevant　biomarkers.