differenti privat support vector machin analysi perceptron dbase activ learn abstract start show activ learn set perceptron algorithm omega ffl label learn linear separ gener error ffl present simpl select sampl algorithm problem combin modif perceptron updat adapt filter rule decid point queri data distribut uniformli unit sphere show algorithm reach gener error ffl log ffl label exponenti improv usual sampl complex supervis learn previous demonstr computation complex queri committe algorithm introduct machin learn applic unlabel data abund labelingi expens distinct captur standard pac onlin model supervis learn motiv field activ learn inwhich label data point initi hidden learner pai label wish reveal queri point chosen randomli numberof label need reach target gener error ffl target confidencelevel ffi similar sampl complex supervis learn hopei altern queri strategi requir significantli fewer privaci dpreserv logist regress paper address import tradeoff privaci learnabl design algorithm learn privat databas focu privaci preserv logist regress appli idea dwork al design privaci preserv logist regress algorithm involv bound sensit regular logist regress perturb learn classifi nois proport sensit show data distribut algorithm poor learn gener compar standard regular logist regress provid privaci preserv regular logist regress algorithm base privaci preserv techniqu solv perturb optim problem prove algorithm preserv privaci model due provid learn guarante show algorithm perform standard regular logist regress term gener error experi demonstr improv learn perform method versu sensit method privaci preserv techniqu depend sensit function extend easili class convex loss function work reveal interest connect regular privaci practic onlin activ learn classif compar practic perform recent propos algorithm activ learn onlin classif set activ learn algorithm combin variant strongli onlin access data sequenti store previous label exampl formal guarante recent proven assumpt motiv optic charact recognit ocr applic argu appropri serv onlin activ learn compar practic efficaci applic algorithm variant show signific reduct label complex random sampl gener agnost activ learn algorithm onlin learn dstationari sequenc onlin learn scenario learner make predict basi fix set expert perform expert chang time manner unknown learner formul class univers learn algorithm problem express simpl bayesian algorithm oper model analog hidden markov model hmm deriv perform bound algorithm consider simpler exist bound bound basi learn rate ident optim expert switch time find analyt express aprioriresolut learn rate paramet extend scalar switch rate result model switch rate aregovern matrix paramet arbitrari homogen hmm appli examin algorithm context problem energi manag resourc alloc sequenti auction 