Improving performance of secure real-number operations
Date
2019-05-07
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Tänapäeval on andmed ja nende analüüsimine laialt levinud ja neist on palju kasu. Selle populaarsuse tõttu on ka rohkem levinud igasugused kombinatsioonid, kuidas andmed ja nende põhjal arvutamine omavahel suhestuda võivad. Meie töö fookuseks on siinkohal need juhtumid, kus andmete omanikud ja need osapooled, kes neid analüüsima peaks, ei lange kas osaliselt või täielikult kokku. Selle näiteks võib tuua meditsiiniandmed, mida nende omanikud tahaks ühest küljest salajas hoida, aga mille kollektiivsel analüüsimine on kasulik. Teiseks näiteks on arvutuste delegeerimine suurema arvutusvõimsusega, ent mitte täiesti usaldusväärsele osapoolele. Valdkond, mis selliseid probleeme uurib, kannab nime turvaline ühisarvutus.
Antud valdkond on eelkõige keskendunud juhtumile, kus andmed on kas täisarvulisel või bitilisel kujul, kuna neid on lihtsam analüüsida ja teised juhtumid saab nendest tuletada, sest kõige, mis üldse arvutatav on, väljaarvutamiseks piisab bittide liitmisest ja korrutamisest. See on teoorias tõsi, samas, kui kõike otse bittide või täisarvude tasemel teha, on tulemus ebaefektiivne. Seepärast vaatleb see doktoritöö turvalist ühisarvutust reaalarvudel ja meetodeid, kuidas seda efektiivsemaks teha.
Esiteks vaatleme ujukoma- ja püsikomaarve. Ujukomaarvud on väga paindlikud ja täpsed, aga on teisalt jälle üsna keeruka struktuuriga. Püsikomaarvud on lihtsa olemusega, ent kannatavad täpsuses. Töö esimene meetod vaatlebki nende kombineerimist, et mõlema häid omadusi ära kasutada.
Teine tehnika baseerub tõigal, et antud paradigmas juhtub, et ei ole erilist ajalist vahet, kas paralleelis teha üks tehe või miljon. Sestap katsume töö teises meetodis teha paralleelselt hästi palju mingit lihtsat operatsiooni, et välja arvutada mõnd keerulisemat.
Kolmas tehnika kasutab reaalarvude kujutamiseks täisarvupaare, (a,b), mis kujutavad reaalarvu a- φb, kus φ=1.618... on kuldlõige. Osutub, et see võimaldab meil üsna efektiivselt liita ja korrutada ja saavutada mõistlik täpsus.
Nowadays data and its analysis are ubiquitous and very useful. Due to this popularity, different combinations of how these two can relate to each other proliferate. We focus on the cases where the owners of the data and those who compute on them don't coincide either partially or totally. Examples are medicinal data where the owners want secrecy but where doing statistics on them collectively is useful, or outsourcing computation. The discipline that studies these cases is called secure computation. This field has been mostly working on integer and bit data types, as they are easier to work on, and due to it being possible to reduce the other cases to integer and bit manipulations. However, using these reductions bluntly will give inefficient results. Thus this thesis studies secure computation on real numbers and presents three methods for improving efficiency. The first method concerns with fixed-point and floating-point numbers. Fixed-point numbers are simple in construction, but can lack precision and flexibility. Floating-point numbers, on the other hand, are precise and flexible, but are rather complicated in nature, which in secure setting translates to expensive operations. The first method thus combines those two number types for greater efficiency. The second method is based on the fact that in the concrete paradigm we use, it does not matter timewise whether we perform one or million operations in parallel. Thus we attempt to perform many instances of a fast operation in parallel in order to evaluate a more complicated one. Thirdly we introduce a new real number type. We use pairs of integers (a,b) to represent the real number a- φb where φ=1.618... is the golden ratio. This number type allows us to perform addition and multiplication relatively quicky and also achieves reasonable granularity.
Nowadays data and its analysis are ubiquitous and very useful. Due to this popularity, different combinations of how these two can relate to each other proliferate. We focus on the cases where the owners of the data and those who compute on them don't coincide either partially or totally. Examples are medicinal data where the owners want secrecy but where doing statistics on them collectively is useful, or outsourcing computation. The discipline that studies these cases is called secure computation. This field has been mostly working on integer and bit data types, as they are easier to work on, and due to it being possible to reduce the other cases to integer and bit manipulations. However, using these reductions bluntly will give inefficient results. Thus this thesis studies secure computation on real numbers and presents three methods for improving efficiency. The first method concerns with fixed-point and floating-point numbers. Fixed-point numbers are simple in construction, but can lack precision and flexibility. Floating-point numbers, on the other hand, are precise and flexible, but are rather complicated in nature, which in secure setting translates to expensive operations. The first method thus combines those two number types for greater efficiency. The second method is based on the fact that in the concrete paradigm we use, it does not matter timewise whether we perform one or million operations in parallel. Thus we attempt to perform many instances of a fast operation in parallel in order to evaluate a more complicated one. Thirdly we introduce a new real number type. We use pairs of integers (a,b) to represent the real number a- φb where φ=1.618... is the golden ratio. This number type allows us to perform addition and multiplication relatively quicky and also achieves reasonable granularity.
Description
Keywords
turvalised ühisarvutused, andmekaitse