I. INTRODUCTIONII. Evaluation parameters and criteriaA. Validated Randomness:1) Birthday Spacings Test:2) Overlapping 5-permutation Test:3) Binary Rank Test for Matrices 31x31 & 32x32:4) Binary Rank Test for 6x8 Matrices:5) Count the Number of 1’s in a stream of bytes:6) Count the Number of 1’s in specific bytes:7) Monkey tests on 20-bit words:8) Monkey tests on OPSO:9) Monkey tests on OQSO:10) Monkey tests on DNA:11) Parking Lot Test:12) Overlapping Sums Test:13) Squeeze Test:14) Minimum Distance Test:15) Random Sphere’s Test:1) Frequency Test:2) Cumulative Sum Test:3) Runs Test:4) Rank Test:5) Spectral Test:6) Templates Matching Test:7) Universal Statistical Test (Maurer):8) Approximate Entropy Test (Pincus & Singer):9) Random Excursions Test:10) Moving Averages Test:11) Lempel-Ziv Compression Test:12) Linear Complexity Test:13) Bayes Test:B. Rated Speed:C. Size:D. Form Factor:E. Self-test Verification:F. Power Requirements:G. Price:H. Operating Temperature Range:III. Market survey and evaluationIV. physical phenomenonV. ConclusionDRAFTAbstract—The goal of this project is to analyze existingcommercial implementations of True Random NumberGenerators and compare the advertised parameters for eachimplementation. Additionally, this paper investigates theprinciples of randomness employed and attempts to classify theimplementations based on their physical attributes andimplementation details.Index Terms—Random number generation, hardware, true I. INTRODUCTIONThere are many applications where random numbers arefundamental to the design of a system. Typical applicationsinclude electronic gambling & lottery applications, advancedgaming applications, and probably most important,Cryptographic key generation. It is obvious that each of theseapplications would cease to be viable if the random numbersthey rely upon could be predicted. Consequently there is acommercially driven need for reliable and unpredictablerandom number generators. In practice, there are two major types of random numbergenerators, pseudo-RNGs, and True-RNGs; both of whichhave been employed in various commercial applications.Pseudo-random number generators are designed usingalgorithms that generate numbers or bit streams that appearto be random. In most cases the output from these RNGs arerandom enough to pass standard statistical testing, but giventhat this method employs a deterministic approach that isinitialized with a seed, it is debatable whether this categorycan be considered genuinely random. True Random NumberGenerators (TRNGs), on the other hand, capitalize onnaturally occurring random phenomena and can generatenearly perfect statistical randomness without being initializedwith seed values. For this reason, system designers shouldstrongly consider the use of TRNGs for any current or futureapplication that depends of randomnessThe purpose of this paper is to investigate the TRNGs that arecommercially available today and to prepare an evaluationthat system designers may refer to when selecting a product.This evaluation is based on performance, randomness testingmethods, physical parameters, and other importantcharacteristics. The information needed for such anevaluation is strictly based on published data from themanufacturers, which is not necessarily consistent between allproducts. Regardless, enough information is publiclyManuscript received May 12, 2005. M. R. Dugan is a computer engineering student at George Mason Universityunder the supervision of Dr. Kris Gaj. ([email protected])available to establish basic criteria upon which a comparisoncan be made. In addition to identifying the comparisoncriteria and forming an evaluation, this paper also identifiesthe different entropy sources for these TRNGs and discussesthe need and different types of post processing. Lastly, thisreport will highlight some of the standardization efforts thatare currently underway for TRNG design.II.EVALUATION PARAMETERS AND CRITERIAThere are several attributes that need to be considered whencomparing True Random Number Generators. In this section,these attributes are described and are used by the author todevelop a prioritized list of design attributes. It should benoted that this prioritized list is a suggested means ofcomparison and should only be used as a guide for systemdesigners, because specific applications may predicate adifferent prioritization. In the sections that follow a briefdescription of each parameter/attribute is given with the mostimportant attributes listed first.A. Validated Randomness:Probably the most important characteristic of a TRNG isthe ability to consistently provide perfect (or nearly perfect)random numbers. Before a TRNG can even be considered itneeds to successfully pass statistical testing. Among theTRNGs surveyed, the two most prominent testing suites citedwere DIEHARD (developed by Dr. George Marsaglia atFlorida State University in the early 1990’s) and the NISTstatistical testing suites for RNGs specified in SpecialPublication 800-22. These testing suites are considered“stringent” because they represent testing that basic RNGsusually fail. Therefore, any TRNGs that does not pass eitherof these testing suites should not be considered. It should alsobe noted that while FIPS 140-1&2 require statistical testingfor random number generation, these standards do not requiretesting in accordance with either DIEHARD or SP 800-22. In search for better random numbers, Dr. Marsaglia noticedthat most “simple” RNGs, which he defined as RNGs withlinear transformation algorithms based on congruential, shift-register, or lagged-Fibbonacci designs, exhibited non-randomcharacteristics when challenged against advanced statisticaltesting. Consequently the DIEHARD test suite was developedand contains fifteen advanced statistical tests each of which isdesigned to produce a “p-value.” All p-value scores are in therange of [0,1] and failures constitute values at the extremeends of the range (within six significant digits 0.000001 or0.999999). The following is a brief