Adversarial　attack　in　the　black-box　scenario　have　become　a　highly　threatening　security　issue　in　natural　language　processing　tasks　such　as　text　classification　and　sentiment　analysis.　Unlike　continuous　images,　text　space　is　discrete,　and　attack　cannot　be　achieved　by　directly　perturbing　the　input,　making　it　challenging　to　generate　adversarial　texts.　The　current　adversarial　attack　methods　are　mainly　based　on　the　heuristic　single-path　strategy　by　finding　the　synonyms　of　some　tokens　in　the　text　and　replacing　them.　However,　these　methods　tend　to　generate　adversarial　examples　with　limited　attack　performance,　weak　semantic　preservation,　and　drastic　fluency　change.　This　paper　innovatively　proposes　an　edit-based　adversarial　text　generation　method　to　trade-off　between　test　accuracy　decrease　and　stealthiness　for　generated　adversarial　texts.　Extensive　experimental　results　demonstrate　that　our　method　outperforms　baselines.　Our　AdvGen　algorithm　achieved　a　2.5%　improvement　in　attack　effectiveness　on　the　IMDB　dataset.　At　the　same　time,　our　method　reduces　the　Levenshtein　distance　by　5.9%,　increases　the　semantic　similarity　by　2.5%　on　the　AG　dataset,　and　has　an　overall　lower　perplexity　on　all　datasets.　©　2024　IEEE.