Research And Implementation Of Collusion Attack And Defense For Peer Review Systems

Peer review evaluates submitted projects in a field by qualified peers.It is widely used in the assignment of reviewers in academic journals and conferences,playing an essential role in ensuring the quality of research results.With the proliferation of resea rch results,various academic journals and conferences use peer-review systems to assi gn reviewers.The assignment results may directly determine the acceptance of researc h results,so studying the security of peer review systems is an urgent and valuable pra ctical problem.The peer review system uses Natural Language Processing(NLP)techniques to m odel the collected papers and reviewers’ expertise and calculate similarity scores.Attac ks at the paper submission stage may lead to serious security problems in the assignme nt.However,existing attacks against NLP models are challenging to migrate directly t o peer review systems.Given the vital role of peer review in the development of acade mic research,this paper investigates peer review systems from the perspective of collu sion attacks and defenses.(1)To address the problem of collusion attack under the peer review system,we f ormally define the problem of collusion against attack in the peer review system,prov e by a derivation that the problem is an NP-hard combinatorial optimization problem,and propose the SIAttack algorithm for word replacement based on synonym dictionar y and semantic similarity.By replacing a small number of words in the target text with synonyms in the colluders’ documents and experimenting with public datasets in the b uilt peer review system environment,two baseline algorithms are designed and compa red from various perspectives such as different perturbation rates,number of iterations and number of colluders to verify that the SIAttack algorithm proposed in this paper c an effectively manipulate the peer review by constructing adversarial samples The assi gnment results of reviewers.(2)This paper finds that the combination of synonym clustering and random syno nym substitution can play a role in disrupting the attack effect of replacing synonyms i n adversarial attacks,studies the defense method against collusion attacks on peer revi ew systems,formally defines the threat model and problem,designs the CARM word mapping algorithm to defend against adversarial attacks on peer review systems,and u ses it in the built peer review system environment.The experiments are conducted in t he public dataset.Three reasonable baselines are constructed based on word-level adv ersarial defense frontier algorithms in other scenarios,demonstrating that the CARM word mapping algorithm has higher accuracy in the peer review system while the algo rithm is effective.(3)In order to visualize the experimental effects of the attack and defense algorith ms,a peer-review collusion attack and defense system is constructed in this paper.Tw o pages of the system setup page and user interaction page are designed by selecting th e appropriate development environment and tools.