Anomaly　detection　on　attributed　networks　has　become　extremely　important　as　attributed　networks　are　modeled　in　various　real-world　scenarios　such　as　financial,　communication,　and　social　networks.　Traditional　anomaly　detection　methods　such　as　degree-based　or　ego-network-based　methods　do　not　address　well　the　complex　interactions　of　attributes　and　structures　in　graph　structures.　Nowadays,　graph　autoencoder-based　methods　are　a　dominant　graph　anomaly　detection　method　as　the　deep　learning　approach　that　better　captures　the　complex　interactions　of　graph　data.　However,　the　current　approach　ignores　how　to　change　the　structure　of　the　graph　to　improve　the　performance　of　graph　anomaly　detection　leading　to　less　than　optimal　anomaly　detection　results　because　there　are　many　invalid　or　even　obstructive　edges　in　the　original　graph,　which　largely　affect　the　results　of　graph　anomaly　detection.　To　address　the　above　problems,　we　propose　an　applicable　graph　update　method　for　graph　anomaly　detection,　and　we　combine　this　method　with　that　of　a　graph　autoencoder　to　construct　a　new　unsupervised　graph　anomaly　detection　framework　to　better　utilize　the　updated　graph　structure.　Specifically,　we　first　make　the　anomalous　nodes　easier　to　distinguish　by　making　similar　nodes　connected　and　removing　unreasonable　edges　through　graph　updating.　Then　we　use　a　graph　autoencoder　to　reconstruct　the　error　on　the　updated　graph.　Meanwhile　the　nodes　with　large　errors　we　consider　to　be　anomalous　nodes.　Particularly,　when　reconstructing　the　error,　we　reconstruct　the　original　node　attributes　through　the　attributes　of　similar　neighbors,　which　can　better　utilize　the　updated　structure　to　improve　the　performance　of　anomaly　detection.　Experimental　results　show　that　our　proposed　framework　outperforms　the　state-of-the-art　baseline　methods　on　all　five　real-world　benchmark　datasets.