Herbert west: deanonymizer

摘要: The vast majority of scientific journal, conference, and grant selection processes withhold the names reviewers from original submitters, taking a better-safe-than-sorry approach for maintaining collegiality within small-world communities academia. While contents review may not color long-term relationship between submitter reviewer, it is best to require us all be saints. This paper raises question whether assumption reviewer anonymity still holds in face readily-available, high-quality machine learning toolkits. Our threat model focuses on how member community might, over time, amass large number unblinded reviews by serving conference committees. We show that with access even relatively small corpus such reviews, simple classification techniques existing toolkits successfully identify reasonably high accuracy. discuss implications findings describe some potential technical policy-based countermeasures.