Death Note: L, Anonymity & Eluding Entropy

Applied Computer Science: On Murder Considered as STEM Field; using information theory to quantify the magnitude of Light Yagami’s mistakes in Death Note and considering fixes
anime, criticism, computer-science, cryptography, Bayes, insight-porn
2011-05-042017-12-15 finished certainty: highly likely importance: 7

In the manga Death Note, the pro­tag­o­nist Light Yagami is given the su­per­nat­ural weapon “Death Note” which can kill any­one on de­mand, and be­gins us­ing it to re­shape the world. The ge­nius de­tec­tive L at­tempts to track him down with analy­sis and trick­ery, and ul­ti­mately suc­ceeds. Death Note is al­most a thought-ex­per­i­men­t-given the per­fect mur­der weapon, how can you screw up any­way? I con­sider the var­i­ous steps of L’s process from the per­spec­tive of com­puter se­cu­ri­ty, cryp­tog­ra­phy, and in­for­ma­tion the­o­ry, to quan­tify Light’s ini­tial anonymity and how L grad­u­ally de-anonymizes him, and con­sider which mis­take was the largest as fol­lows:

  1. Light’s fun­da­men­tal mis­take is to kill in ways un­re­lated to his goal.

    Killing through heart at­tacks does not just make him vis­i­ble early on, but the deaths re­veals that his as­sas­si­na­tion method is im­pos­si­bly pre­cise and some­thing pro­foundly anom­alous is go­ing on. L has been tipped off that Kira ex­ists. What­ever the bo­gus jus­ti­fi­ca­tion may be, this is a ma­jor vic­tory for his op­po­nents. (To de­ter crim­i­nals and vil­lains, it is not nec­es­sary for there to be a glob­al­ly-known sin­gle anom­alous or su­per­nat­ural killer, when it would be equally effec­tive to arrange for all the killings to be done nat­u­ral­is­ti­cally by or­di­nary mech­a­nisms such as third par­ties/po­lice/ju­di­ciary or used in­di­rectly as par­al­lel con­struc­tion to crack cas­es.)

  2. Worse, the deaths are non-ran­dom in other ways—they tend to oc­cur at par­tic­u­lar times!

    Just the sched­ul­ing of deaths cost Light 6 bits of anonymity

  3. Light’s third mis­take was re­act­ing to the bla­tant provo­ca­tion of Lind L. Tai­lor.

Tak­ing the bait let L nar­row his tar­get down to 1⁄3 the orig­i­nal Japan­ese pop­u­la­tion, for a gain of ~1.6 bits. 4. Light’s fourth mis­take was to use con­fi­den­tial po­lice in­for­ma­tion stolen us­ing his po­lice­man fa­ther’s cre­den­tials.

This mis­take was the largest in bits lost. This mis­take cost him 11 bits of anonymi­ty; in other words, this mis­take cost him twice what his sched­ul­ing cost him and al­most 8 times the mur­der of Tai­lor! 5. Killing Ray Pen­bar and the FBI team.

If we as­sume Pen­bar was tasked 200 leads out of the 10,000, then mur­der­ing him and the fi­ancee dropped Light just 6 bits or a lit­tle over half the fourth mis­take and com­pa­ra­ble to the orig­i­nal sched­ul­ing mis­take. 6. Endgame: At this point in the plot, L re­sorts to di­rect mea­sures and en­ters Light’s life di­rect­ly, en­rolling at the uni­ver­si­ty, with Light un­able to per­fectly play the role of in­no­cent un­der in­tense in­-per­son sur­veil­lance.

From that point on, Light is screwed as he is now play­ing a deadly game of “Mafia” with L & the in­ves­tiga­tive team. He frit­tered away >25 bits of anonymity and then L in­tu­ited the rest and sus­pected him all along.

Fi­nal­ly, I sug­gest how Light could have most effec­tively em­ployed the Death Note and lim­ited his loss of anonymi­ty. In an ap­pen­dix, I dis­cuss the max­i­mum amount of in­for­ma­tion leak­age pos­si­ble from us­ing a Death Note as a com­mu­ni­ca­tion de­vice.

(Note: This es­say as­sumes a fa­mil­iar­ity with the early plot of and . If you are un­fa­mil­iar with DN, see my es­say or con­sult or read the DN rules.)

I have called the pro­tag­o­nist of Death Note, Light Yagami, “hubris­tic” and said he made big mis­takes. So I ought to ex­plain what he did wrong and how he could do bet­ter.

While Light starts schem­ing and tak­ing se­ri­ous risks as early as the ar­rival of the FBI team in Japan, he has fun­da­men­tally al­ready screwed up. L should never have got­ten that close to Light. The Death Note kills flaw­lessly with­out foren­sic trace and over ar­bi­trary dis­tances; Death Note is al­most a thought-ex­per­i­men­t—­given the per­fect mur­der weapon, how can you screw up any­way?

Some of the other Death Note users high­light the prob­lem. The user in the car­ries out the nor­mal ex­e­cu­tions, but also kills a num­ber of promi­nent com­peti­tors. The killings di­rectly point to the Yot­suba Group and even­tu­ally the user’s death. The moral of the story is that in­di­rect re­la­tion­ships can be fa­tal in nar­row­ing down the pos­si­bil­i­ties from ‘every­one’ to ‘these 8 men’.

Detective stories as optimization problems

In Light’s case, L starts with the world’s en­tire pop­u­la­tion of 7 bil­lion peo­ple and needs to nar­row it down to 1 per­son. It’s a search prob­lem. It maps fairly di­rectly onto ba­sic , in fact. (See also , , and for case stud­ies in ap­plied deanonymiza­tion, .) To uniquely spec­ify one item out of 7 bil­lion, you need 33 bits of in­for­ma­tion be­cause log2(7000000000) ≈ 32.7; to use an anal­o­gy, your 32-bit com­puter can only ad­dress one unique lo­ca­tion in mem­ory out of 4 bil­lion lo­ca­tions, and adding an­other bit dou­bles the ca­pac­ity to >8 bil­lion. Is 33 bits of in­for­ma­tion a lot?

Not re­al­ly. L could get one bit just by look­ing at his­tory or crime sta­tis­tics, and not­ing that mass mur­der­ers are, to an as­ton­ish­ing de­gree, male1, thereby rul­ing out half the world pop­u­la­tion and ac­tu­ally start­ing L off with a re­quire­ment to ob­tain only 32 bits to break Light’s anonymi­ty.2 If Death Note users were suffi­ciently ra­tio­nal & knowl­edge­able, they could draw on con­cepts like to acausally co­op­er­ate3 to avoid this in­for­ma­tion leak­age… by ar­rang­ing to pass on Death Notes to fe­males4 to re­store a 50:50 gen­der ra­tio—­for ex­am­ple, if for every fe­male who ob­tained a Death note there were 3 males with Death Notes, then all users could roll a 1d3 dice and if 1 keep it and if 2 or 3 pass it on to some­one of the op­po­site gen­der.

We should first point out that Light is al­ways go­ing to leak some bits. The only way he could re­main per­fectly hid­den is to not use the Death Note at all. If you change the world in even the slight­est way, then you have leaked in­for­ma­tion about your­self in prin­ci­ple. Every­thing is con­nected in some sense; you can­not mag­i­cally wave away the ex­is­tence of fire with­out cre­at­ing a cas­cade of con­se­quences that re­sult in every liv­ing thing dy­ing. For ex­am­ple, the fun­da­men­tal point of Light ex­e­cut­ing crim­i­nals is to shorten their lifes­pan—there’s no way to hide that. You can’t both shorten their lives and not shorten their lives. He is go­ing to re­veal him­self this way, at the least, to the ac­tu­ar­ies and sta­tis­ti­cians.

More his­tor­i­cal­ly, this has been a chal­lenge for cryp­tog­ra­phers, like in WWII: how did they ex­ploit the Enigma & other com­mu­ni­ca­tions with­out re­veal­ing they had done so? Their so­lu­tion was mis­di­rec­tion: , like search planes that ‘just hap­pened’ to find Ger­man sub­marines or leaks to about there be­ing undis­cov­ered spies. (How­ev­er, the fa­mous story that Win­ston Churchill al­lowed the town of Coven­try to be bombed rather than risk the se­cret of Ul­tra has .) This worked in part be­cause of Ger­man over­con­fi­dence, be­cause the war did not last too long, and in part be­cause each cover story was plau­si­ble on its own and no one was, in the chaos of war, able to see the whole pic­ture and re­al­ize that there were too many lucky search planes and too many undis­cov­er­able moles; even­tu­al­ly, how­ev­er, some­one would re­al­ize, and ap­par­ently some Ger­mans did con­clude that Enigma had to have been bro­ken (but much too late). It’s not clear to me what would be the best mis­di­rec­tion for Light to mask his nor­mal killings—use the Death Note’s con­trol fea­tures to in­vent a an­ti-crim­i­nal ter­ror­ist or­ga­ni­za­tion?

So there is a real chal­lenge here: one party is try­ing to in­fer as much as pos­si­ble from ob­served effects, and the other is try­ing to min­i­mize how much the for­mer can ob­serve while not stop­ping en­tire­ly. How well does Light bal­ance the com­pet­ing de­mands?


Mistake 1

How­ev­er, he can try to re­duce the leak­age and make his as large as pos­si­ble. For ex­am­ple, killing every crim­i­nal with a heart at­tack is a dead give-away. Crim­i­nals do not die of heart at­tacks that often. (The point is more dra­matic if you re­place ‘heart at­tack’ with ‘lu­pus’; as we all know, in real life it’s never lu­pus.) Heart at­tacks are a sub­set of all deaths, and by re­strict­ing him­self, Light makes it eas­ier to de­tect his ac­tiv­i­ties. 1000 deaths of lu­pus are a blar­ing red alarm; 1000 deaths of heart at­tacks are an odd­i­ty; and 1000 deaths dis­trib­uted over the sta­tis­ti­cally likely sus­pects of can­cer and heart dis­ease etc. are al­most in­vis­i­ble (but still no­tice­able in prin­ci­ple).

So, Light’s fun­da­men­tal mis­take is to kill in ways un­re­lated to his goal. Killing through heart at­tacks does not just make him vis­i­ble early on, but the deaths re­veals that his as­sas­si­na­tion method is su­per­nat­u­rally pre­cise. L has been tipped off that Kira ex­ists. What­ever the bo­gus jus­ti­fi­ca­tion may be, this is a ma­jor vic­tory for his op­po­nents.

First mis­take, and a clas­sic one of se­r­ial killers (eg the ’s vaunt­ing was less anony­mous than he be­lieved): delu­sions of grandeur and the de­sire to taunt, play with, and con­trol their vic­tims and demon­strate their power over the gen­eral pop­u­la­tion. From a lit­er­ary per­spec­tive, this sim­i­lar­ity is clearly not an ac­ci­dent, as we are meant to read Light as the So­ciopath Hero ar­che­type: his ul­ti­mate down­fall is the con­se­quence of his , , par­tic­u­larly in the orig­i­nal sadis­tic sense. Light can­not help but self­-s­ab­o­tage like this.

(This is also deeply prob­lem­atic from the point of car­ry­ing out Light’s the­ory of de­ter­rence: to de­ter crim­i­nals and vil­lains, it is not nec­es­sary for there to be a glob­al­ly-known sin­gle su­per­nat­ural killer, when it would be equally effec­tive to arrange for all the killings to be done nat­u­ral­is­ti­cally by third par­ties/po­lice/ju­di­ciary or used in­di­rectly to crack cas­es. Ar­guably the de­ter­rence would be more effec­tive the more diffused it’s be­lieved to be—s­ince a sin­gle killer has a fi­nite lifes­pan, fi­nite knowl­edge, fal­li­bil­i­ty, and idio­syn­cratic pref­er­ences which re­duce the threat and con­nec­tion to crim­i­nal­i­ty, while if all the deaths were as­cribed to un­usu­ally effec­tive po­lice or de­tec­tives, this would be in­ferred as a gen­eral in­crease in all kinds of po­lice com­pe­tence, one which will not in­stantly dis­ap­pear when one per­son gets bored or hit by a bus.)

Mistake 2

Worse, the deaths are non-ran­dom in other ways—they tend to oc­cur at par­tic­u­lar times! Graphed, daily pat­terns jump out.

L was able to nar­row down the ac­tive times of the pre­sum­able stu­dent or worker to a par­tic­u­lar range of lon­gi­tude, say 125–150° out of 180°; and what coun­try is most promi­nent in that range? Japan. So that cut down the 7 bil­lion peo­ple to around 0.128 bil­lion; 0.128 bil­lion re­quires 27 bits (log2 (128000000) ≈ 26.93) so just the sched­ul­ing of deaths cost Light 6 bits of anonymi­ty!


On a side-note, some might be skep­ti­cal that one can in­fer much of any­thing from the graph and that Death Note was just gloss­ing over this part. “How can any­one in­fer that it was some­one liv­ing in Japan just from 2 clumpy lines at morn­ing and evening in Japan?” But ac­tu­al­ly, such a graph is sur­pris­ingly pre­cise. I learned this years be­fore I watched Death Note, when I was heav­ily ac­tive on Wikipedia; often I would won­der if two ed­i­tors were the same per­son or roughly where an ed­i­tor lived. What I would do if their ed­its or user page did not re­veal any­thing use­ful is I would go to “Kate’s edit counter” and I would ex­am­ine the times of day all their hun­dreds or thou­sands of ed­its were made at. Typ­i­cal­ly, what one would see was ~4 hours where there were no ed­its what­so­ev­er, then ~4 hours with mod­er­ate to high ac­tiv­i­ty, a trough, then an­other grad­ual rise to 8 hours later and a fur­ther de­cline down to the first 4 hours of no ac­tiv­i­ty. These pe­ri­ods quite clearly cor­re­sponded to sleep (pretty much every­one is asleep at 4 AM), morn­ing, lunch & work hours, evening, and then night with peo­ple oc­ca­sion­ally stay­ing up late and edit­ing5. There was noise, of course, from peo­ple stay­ing up es­pe­cially late or get­ting in a bunch of edit­ing dur­ing their work­day or oc­ca­sion­ally trav­el­ing, but the over­all pat­terns were clear—n­ever did I dis­cover that some­one was ac­tu­ally a night­watch­man and my guess was an en­tire hemi­sphere off. (A­ca­d­e­mic es­ti­mates based on user edit­ing pat­terns cor­re­late well with what is pre­dicted by on the ba­sis of the ge­og­ra­phy of IP ed­its.6)

Com­puter se­cu­rity re­search offers more scary re­sults. Per­haps be­cause , there are an amaz­ing num­ber of ways to break some­one’s pri­vacy and de-anonymize them (back­ground; there is also fi­nan­cial in­cen­tive to do so in or­der to ad­ver­tise & ):

  1. small er­rors in their com­put­er’s clock’s time (even over Tor)

  2. Web brows­ing his­tory7 or just the ver­sion and plu­g­ins8; and this is when ran­dom Fire­fox or Google Docs or Face­book bugs don’t leak your iden­tity

  3. based on how slow pages load9 (how many there are; tim­ing at­tacks can also be used to learn web­site user­names or # of pri­vate pho­tos)

  4. Knowl­edge of what ‘groups’ a per­son was in could uniquely iden­tify 42%10 of peo­ple on so­cial net­work­ing site , and pos­si­bly Face­book & 6 oth­ers

  5. Sim­i­lar­ly, some­one has watched11, pop­u­lar or ob­scure, through often grants ac­cess to the rest of their pro­file if it was in­cluded in the . (This was more dra­matic than the be­cause AOL searches had a great deal of per­sonal in­for­ma­tion em­bed­ded in the search queries, but in con­trast, the Net­flix data seems im­pos­si­bly im­pov­er­ished—there’s noth­ing ob­vi­ously iden­ti­fy­ing about what anime one has watched un­less one watches ob­scure ones.)

  6. The re­searchers to find iso­mor­phisms be­tween ar­bi­trary graphs12 (such as so­cial net­works stripped of any and all data ex­cept for the graph struc­ture), for ex­am­ple and , and give many ex­am­ples of pub­lic datasets that could be de-anonymized13—such as your Ama­zon pur­chases (Ca­lan­drino et al 2011; blog). These at­tacks are on just the data that is left after at­tempts to anonymize data; they don’t ex­ploit the ob­ser­va­tion that the choice of what data to re­move is as in­ter­est­ing as what is left, what calls “The Redac­tor’s Dilemma”.

  7. User­names hardly bear dis­cussing

  8. Your hos­pi­tal records can be de-anonymized just by look­ing at pub­lic vot­ing rolls14 That re­searcher later went on to run “ex­per­i­ments on the iden­ti­fi­a­bil­ity of de-i­den­ti­fied sur­vey data [cite], phar­macy data [cite], clin­i­cal trial data [cite], crim­i­nal data [S­tate of Delaware v. Gan­nett Pub­lish­ing], DNA [cite, cite, cite], tax data, pub­lic health reg­istries [cite (sealed by court), etc.], web logs, and par­tial So­cial Se­cu­rity num­bers [cite].” (Whew.)

  9. Your is sur­pris­ingly unique and the sounds of typ­ing and arm move­ments can iden­tify you or be used snoop on in­put &

  10. Know­ing your morn­ing com­mute as loosely as to the in­di­vid­ual blocks (or less gran­u­lar) uniquely iden­ti­fies (Golle & Par­tridge 2009) you; know­ing your com­mute to the zip code/­cen­sus tract uniquely iden­ti­fies 5% of peo­ple

  11. Your hand­writ­ing is fairly unique, sure—but so is how you fill in bub­bles on tests15

  12. Speak­ing of hand­writ­ing, your writ­ing style can be pretty unique too

  13. the un­no­tice­able back­ground elec­tri­cal hum may uniquely date au­dio record­ings. Un­no­tice­able sounds can also be used to per­sis­tently track de­vices/peo­ple, ex­fil­trate in­for­ma­tion across air gaps, and can be used to mon­i­tor room pres­ence/ac­tiv­i­ty, and even or tap­ping noises or

  14. you may have heard of for eaves­drop­ping… but what about eaves­drop­ping via video record­ing of potato chip bags or candy wrap­pers or ? (press re­lease), or cell­phone gy­ro­scopes? Lasers are good for de­tect­ing your heart­beat as well, which is—of course—uniquely iden­ti­fy­ing And Soon even will no longer be safe…

  15. steer­ing & dri­ving pat­terns are suffi­ciently unique as to al­low iden­ti­fi­ca­tion of dri­vers from as lit­tle as 1 turn in some cas­es: . These at­tacks also work on smart­phones for time zone, baro­met­ric pres­sure, pub­lic trans­porta­tion tim­ing, IP ad­dress, & pat­tern of con­nect­ing to WiFi or cel­lu­lar net­works (Mose­nia et al 2017)

  16. smart­phones can be IDed by the pat­tern of pixel noise, due to sen­sor noise such as small im­per­fec­tions in the CCD sen­sors and lenses (and Face­book has even patented this)

  17. smart­phone us­age pat­terns, such as app pref­er­ences, app switch­ing rates, con­sis­tency of com­mute pat­terns, over­all ge­o­graphic mo­bil­i­ty, slower or less dri­ving have been cor­re­lated with Alzheimer’s dis­ease (Kour­tis et al 2019) and per­son­al­ity ().16

    Eye track­ing is .

  18. voices cor­re­late with not just age/­gen­der/eth­nic­i­ty, but… ?

(The only sur­pris­ing thing about DNA-related pri­vacy breaks is how long they have taken to show up.)

To sum­ma­rize: is al­most im­pos­si­ble17 and pri­vacy is dead18. (See also “Bro­ken Promises of Pri­va­cy: Re­spond­ing to the Sur­pris­ing Fail­ure of Anonymiza­tion”.)

Mistake 3

Light’s third mis­take was re­act­ing to the provo­ca­tion of the Lind L. Tai­lor broad­cast, crit­i­ciz­ing Ki­ra, and Light lash­ing out to use the clear­ly-vis­i­ble name & face to kill Lind L. Tai­lor. The live broad­cast was a bla­tant at­tempt to pro­voke a re­ac­tion—any re­ac­tion—from a sur­prised & un­pre­pared Light, and that alone should have been suffi­cient rea­son to sim­ply ig­nore it (even if Light could not have rea­son­ably known ex­actly how it was a trap): one should never do what an en­emy wants one to do on ground & terms & tim­ing pre­pared by the en­e­my. (Light had the op­tion to use the Death Note at any time in the fu­ture, and that would have been al­most as good a demon­stra­tion of his power as do­ing so dur­ing a live broad­cast.)

Run­ning the broad­cast in 1 re­gion was also a gam­ble & a po­ten­tial mis­take on L’s part; he had no real rea­son to think Light was in (or if he did al­ready have pri­ors/in­for­ma­tion to that effect, he should’ve been bi­sect­ing Kan­to) and should have arranged for it to be broad­cast to ex­actly half of Japan’s pop­u­la­tion, ob­tain­ing an ex­pected max­i­mum of 1 bit. But it was one that paid off; he nar­rowed his tar­get down to 1⁄3 the orig­i­nal Japan­ese pop­u­la­tion, for a gain of ~1.6 bits. (You can see it was a gam­ble by con­sid­er­ing if Light had been out­side Kan­to; since he would not see it live, he would not have re­act­ed, and all L would learn is that his sus­pect was in that other 2⁄3 of the pop­u­la­tion, for a gain of only ~0.3 bit­s.)

But even this was­n’t a huge mis­take. He lost 6 bits to his sched­ule of killing, and lost an­other 1.6 bits to tem­pera­men­tally killing Lind L. Tai­lor, but since the male pop­u­la­tion of Kanto is 21.5 mil­lion (43 mil­lion to­tal), he still has ~24 bits of anonymity left (log2 (21500000) ≈ 24.36). That’s not too ter­ri­ble, and the loss is mit­i­gated even fur­ther by other de­tails of this mis­take, as pointed out by Zm­flav­ius; specifi­cal­ly, that un­like “be­ing male” or “be­ing Japan­ese”, the in­for­ma­tion about be­ing in Kanto is sub­ject to de­cay, since peo­ple move around all the time for all sorts of rea­sons:

…quite pos­si­bly Light’s biggest mis­take was in­ad­ver­tently re­veal­ing his con­nec­tion to the po­lice hi­er­ar­chy by hack­ing his dad’s com­put­er. Whereas even the Lind L. Tay­lor de­ba­cle only re­vealed his killing me­chan­ics and nar­rowed him down to “some­one in the Kanto re­gion” (which is, while an im­pres­sive ac­com­plish­ment based on the in­for­ma­tion he had, en­tirely mean­ing­less for ac­tu­ally find­ing a sus­pec­t), there were per­haps a few hun­dred peo­ple who had ac­cess to the in­for­ma­tion Light’s dad had. There’s also the fact that L knew that Light was prob­a­bly some­one in their late teens, mean­ing that there was an ex­tremely high chance that at the end of the school year, even that coup of his would ex­pire, thanks to stu­dents head­ing off to uni­ver­sity all over Japan (of course, Light went to , and a stu­dent of his cal­iber not at­tend­ing such a uni­ver­sity would be sus­pi­cious, but L had no way of know­ing that then). I mean, per­haps L had hoped that Kira would re­veal him­self by sud­denly mov­ing away from the Kanto re­gion, but come the next May, he would have no way of mon­i­tor­ing un­usual move­ments among late teenagers, be­cause a large per­cent­age of them would be mov­ing for le­git­i­mate rea­sons.

(One could still run the in­fer­ence “back­wards” on any par­tic­u­lar per­son to ver­ify they were in Kanto in the right time pe­ri­od, but as time pass­es, it be­comes less pos­si­ble to run the in­fer­ence “for­wards” and only ex­am­ine peo­ple in Kan­to.)

This mis­take also shows us that the im­por­tant thing that in­for­ma­tion the­ory buys us, re­al­ly, is not the bit (we could be us­ing log10 rather than log2, and com­pares rather than “bits”) so much as com­par­ing events in the plot on a log­a­rith­mic scale. If we sim­ply looked at how the ab­solute num­ber of how many peo­ple were ruled out at each step, we’d con­clude that the first mis­take by Light was a de­ba­cle with­out com­pare since it let L rule out >6 bil­lion peo­ple, ap­prox­i­mately 60× more peo­ple than all the other mis­takes put to­gether would let L rule out. Mis­takes are rel­a­tive to each oth­er, not ab­solutes.

Mistake 4

Light’s fourth mis­take was to use con­fi­den­tial po­lice in­for­ma­tion stolen us­ing his po­lice­man fa­ther’s cre­den­tials. This was un­nec­es­sary as there are count­less crim­i­nals he could still ex­e­cute us­ing pub­lic in­for­ma­tion (face+­name is not typ­i­cally diffi­cult to get), and if for some rea­son he needed a spe­cific crim­i­nal, he could ei­ther re­strict use of se­cret in­for­ma­tion to a few high­-pri­or­ity vic­tim­s—if only to avoid sus­pi­cions of hack­ing & sub­se­quent se­cu­rity up­grades cost­ing him ac­cess!—or man­u­fac­ture, us­ing the Death Note’s co­er­cive pow­ers or Ki­ra’s pub­lic sup­port, a way to re­lease in­for­ma­tion such as a ‘leak’ or pass­ing pub­lic trans­parency laws.

This mis­take was the largest in bits lost. But in­ter­est­ing­ly, many or even most Death Note fans do not seem to re­gard this as his largest mis­take, in­stead point­ing to his killing Lind L. Tai­lor or per­haps re­ly­ing too much on Mika­mi. The in­for­ma­tion the­o­ret­i­cal per­spec­tive strongly dis­agrees, and lets us quan­tify how large this mis­take was.

When he acts on the se­cret po­lice in­for­ma­tion, he in­stantly cuts down his pos­si­ble iden­tity to one out of a few thou­sand peo­ple con­nected to the po­lice. Let’s be gen­er­ous and say 10,000. It takes 14 bits to spec­ify 1 per­son out of 10,000 (log2 (10000) ≈ 13.29)—as com­pared to the 24–25 bits to spec­ify a Kanto dweller.

This mis­take cost him 11 bits of anonymi­ty; in other words, this mis­take cost him twice what his sched­ul­ing cost him and al­most 8 times the mur­der of Tai­lor!

Mistake 5

In com­par­ison, the fifth mis­take, mur­der­ing Ray Pen­bar’s fi­ancee and fo­cus­ing L’s sus­pi­cion on Pen­bar’s as­signed tar­gets was pos­i­tively cheap. If we as­sume Pen­bar was tasked 200 leads out of the 10,000, then mur­der­ing him and the fi­ancee dropped Light from 14 bits to 8 bits (log2 (200) ≈ 7.64) or just 6 bits or a lit­tle over half the fourth mis­take and com­pa­ra­ble to the orig­i­nal sched­ul­ing mis­take.


At this point in the plot, L re­sorts to di­rect mea­sures and en­ters Light’s life di­rect­ly, en­rolling at the uni­ver­si­ty. From this point on, Light is screwed as he is now play­ing a deadly game of with L & the in­ves­tiga­tive team. He frit­tered away >25 bits of anonymity and then L in­tu­ited the rest and sus­pected him all along. (We could jus­tify L skip­ping over the re­main­ing 8 bits by point­ing out that L can an­a­lyze the deaths and in­fer psy­cho­log­i­cal char­ac­ter­is­tics like ar­ro­gance, puz­zle-solv­ing, and great in­tel­li­gence, which com­bined with heuris­ti­cally search­ing the re­main­ing can­di­dates, could lead him to zero in on Light.)

From the the­o­ret­i­cal point of view, the game was over at that point. The chal­lenge for L then be­came prov­ing it to L’s sat­is­fac­tion un­der his self­-im­posed moral con­straints.19

Security is Hard (Let’s Go Shopping)

What should Light have done? That’s easy to an­swer, but tricky to im­ple­ment.

One could try to man­u­fac­ture disin­for­ma­tion. re­hearses many of the above points about in­for­ma­tion the­ory & anonymi­ty, and goes on to loosely dis­cuss the pos­si­ble ben­e­fits of fak­ing in­for­ma­tion:

…one ad­di­tional way to gain more anonymity is through de­lib­er­ate dis­in­for­ma­tion. For in­stance, sup­pose that one re­veals 100 in­de­pen­dent bits of in­for­ma­tion about one­self. Or­di­nar­i­ly, this would cost 100 bits of anonymity (as­sum­ing that each bit was a pri­ori equally likely to be true or false), by cut­ting the num­ber of pos­si­bil­i­ties down by a fac­tor of 2100; but if 5 of these 100 bits (cho­sen ran­domly and not re­vealed in ad­vance) are de­lib­er­ately fal­si­fied, then the num­ber of pos­si­bil­i­ties in­creases again by a fac­tor of (100 choose 5) ~ 226, re­cov­er­ing about 26 bits of anonymi­ty. In prac­tice one gains even more anonymity than this, be­cause to dis­pel the dis­in­for­ma­tion one needs to solve a prob­lem, which can be no­to­ri­ously in­tractable com­pu­ta­tion­al­ly, al­though this ad­di­tional pro­tec­tion may dis­si­pate with time as al­go­rithms im­prove (e.g. by in­cor­po­rat­ing ideas from ).


The diffi­culty with sug­gest­ing that Light should—or could—have used dis­in­for­ma­tion on the tim­ing of deaths is that we are, in effect, en­gag­ing in a sort of . How ex­actly is Light or any­one sup­posed to know that L could de­duce his time­zone from his killings? I men­tioned an ex­am­ple of us­ing Wikipedia ed­its to lo­cal­ize ed­i­tors, but that tech­nique was unique to me among WP ed­i­tors20 and no doubt there are many other forms of in­for­ma­tion leak­age I have never heard of de­spite com­pil­ing a list; if I were Light, even if I re­mem­bered my Wikipedia tech­nique, I might not bother evenly dis­trib­ut­ing my killing over the clock or adopt­ing a de­cep­tive pat­tern (eg sug­gest­ing I was in Eu­rope rather than Japan). If Light had known he was leak­ing tim­ing in­for­ma­tion but did­n’t know that some­one out there was clever enough to use it (a “known un­known”), then we might blame him; but how is Light sup­posed to know these “un­known un­knowns”?

is the an­swer. Ran­dom­iza­tion and en­cryp­tion scram­ble the cor­re­la­tions be­tween in­put and out­put, and they would serve as well in Death Note as they do in cryp­tog­ra­phy & sta­tis­tics in the real world, at the cost of some effi­cien­cy. The point of ran­dom­iza­tion, both in cryp­tog­ra­phy and in sta­tis­ti­cal ex­per­i­ments, is to not just pre­vent the leaked in­for­ma­tion or (re­spec­tive­ly) you do know about but also the ones you do not yet know about.

To steal & para­phrase an ex­am­ple from Un­con­trolled: you’re run­ning a weight-loss ex­per­i­ment. You know that the effec­tive­ness might vary with each sub­jec­t’s pre-ex­ist­ing weight, but you don’t be­lieve in ran­dom­iza­tion (y­ou’re a prac­ti­cal man! only prissy sta­tis­ti­cians worry about ran­dom­iza­tion!); so you split the sub­jects by weight, and for con­ve­nience you al­lo­cate them by when they show up to your ex­per­i­men­t—in the end, there are ex­actly 10 ex­per­i­men­tal sub­jects over 150 pounds and 10 con­trols over 150 pounds, and so on and so forth. Un­for­tu­nate­ly, it turns out that un­be­knownst to you, a ge­netic vari­ant con­trols weight gain and a whole ex­tended fam­ily showed up at your ex­per­i­ment early on and they all got al­lo­cated to ‘ex­per­i­men­tal’ and none of them to ‘con­trol’ (s­ince you did­n’t need to ran­dom­ize, right? you were mak­ing sure the groups were matched on weight!). Your ex­per­i­ment is now bo­gus and mis­lead­ing. Of course, you could run a sec­ond ex­per­i­ment where you make sure the ex­per­i­men­tal and con­trol groups are matched on weight and also now matched on that ge­netic vari­ant… but now there’s the po­ten­tial for some third con­founder to hit you. If only you had used ran­dom­iza­tion—then you would prob­a­bly have put some of the vari­ants into the other group as well and your re­sults would­n’t’ve been bo­gus!

So to deal with Light’s first mis­take, sim­ply sched­ul­ing every death on the hour will not work be­cause the wake-sleep cy­cle is still pre­sent. If he set up a list and wrote down n crim­i­nals for each hour to elim­i­nate the peak-troughs rather than ran­dom­iz­ing, could that still go wrong? May­be: we don’t know what in­for­ma­tion might be left in the data which an L or Tur­ing could de­ci­pher. I can spec­u­late about one pos­si­bil­i­ty—the al­lo­ca­tion of each kind of crim­i­nal to each hour. If one were to draw up lists and go in or­der (hey, one does­n’t need ran­dom­iza­tion, right?), then the or­der might go ‘crim­i­nals in the morn­ing news­pa­per, crim­i­nals on TV, crim­i­nals whose de­tails were not im­me­di­ately given but were avail­able on­line, crim­i­nals from years ago, his­tor­i­cal crim­i­nals etc’; if the morn­ing-news­pa­per-crim­i­nals start at say 6 AM Japan time… And al­lo­cat­ing evenly might be hard, since there’s nat­u­rally go­ing to be short­falls when there just aren’t many crim­i­nals that day or the news­pa­pers aren’t pub­lish­ing (hol­i­days?) etc., so the short­fall pe­ri­ods will pin­point what the Kira con­sid­ers ‘end of the day’.

A much safer pro­ce­dure is thor­ough-go­ing ran­dom­iza­tion ap­plied to tim­ing, sub­jects, and man­ner of death. Even if we as­sume that Light was bound and de­ter­mined to re­veal the ex­is­tence of Kira and gain pub­lic­ity and in­ter­na­tional no­to­ri­ety (a ma­jor char­ac­ter flaw in its own right; ac­com­plish­ing things, tak­ing cred­it—­choose one), he still did not have to re­duce his anonymity much past 32 bits.

  1. Each ex­e­cu­tion’s time could be de­ter­mined by a ran­dom dice roll (say, a 24-sided dice for hours and a 60-sided dice for min­utes).
  2. Se­lect­ing method of death could be done sim­i­larly based on eas­ily re­searched de­mo­graphic data, al­though per­haps ir­rel­e­vant (serv­ing mostly to con­ceal that a killing has taken place).
  3. Se­lect­ing crim­i­nals could be based on in­ter­na­tion­ally ac­ces­si­ble pe­ri­od­i­cals that plau­si­bly every hu­man has ac­cess to, such as the New York Times, and deaths could be de­layed by months or years to broaden the pos­si­bil­i­ties as to where the Kira learned of the vic­tim (TV? books? the In­ter­net?) and avoid­ing is­sues like killing a crim­i­nal only pub­li­cized on one ob­scure Japan­ese pub­lic tele­vi­sion chan­nel. And so on.

Let’s re­mem­ber that all this is pred­i­cated on anonymi­ty, and on Light us­ing low-tech strate­gies; as one per­son asked me, “why does­n’t Light set up an cryp­to­graphic or just take over the world? He would win with­out all this clev­er­ness.” Well, then it would not be Death Note.

See Also


Communicating with a Death Note

One might won­der how much in­for­ma­tion one could send in­ten­tion­ally with a Death Note, as op­posed to in­ad­ver­tently leak bits about one’s iden­ti­ty. As deaths are by and large pub­licly known in­for­ma­tion, we’ll as­sume the sender and re­cip­i­ent have some sort of pre-arranged key or one-time pad (although one would won­der why they’d use such an im­moral and clumsy sys­tem as op­posed to steganog­ra­phy or mes­sages on­line).

A death in­flicted by a Death Note has 3 main dis­tin­guish­ing traits which one can con­trol—who, when, and how:

  1. the per­son

    The ‘who?’ is al­ready cal­cu­lated for us: if it takes 33 bits to spec­ify a unique hu­man, then a par­tic­u­lar hu­man can con­vey 33 bits. Con­cerns about learn­abil­ity (how would you learn of an Ama­zon tribesman’s death?) im­ply that it’s re­ally <33 bits.

    If you try some scheme to en­code more bits into the choice of as­sas­si­na­tion, you ei­ther wind up with 33 bits or you wind up un­able to con­vey cer­tain com­bi­na­tions of bits and effec­tively 33 bits any­way—y­our scheme will tell you that to con­vey your des­per­ately im­por­tant mes­sage X of 50 bits telling all about L’s true iden­tity and how you dis­cov­ered it, you need to kill an Ola­fur Ja­cobs of Tan­za­nia who weighs more than 200 pounds and is from Tai­wan, but alas! Ja­cobs does­n’t ex­ist for you to kill.

  2. the time

    The ‘when’ is han­dled by sim­i­lar rea­son­ing. There is a cer­tain gran­u­lar­ity to Death Note kills: even if it is ca­pa­ble of tim­ing deaths down to the nanosec­ond, one can’t ac­tu­ally wit­ness this or re­ceive records of this. Doc­tors may note time of death down to the min­ute, but no finer (and how do you get such pre­cise med­ical records any­way?). News re­ports may be even less ac­cu­rate, not­ing merely that it hap­pened in the morn­ing or in the late evening. In rare cases like live broad­casts, one may be able to do a lit­tle bet­ter, but even they tend to be de­layed by a few sec­onds or min­utes to al­low for buffer­ing, tech­ni­cal glitches be fixed, the stenog­ra­phers pro­duce the closed cap­tion­ing, or sim­ply to guard against em­bar­rass­ing events (like Janet Jack­son’s nip­ple-s­lip). So we’ll not as­sume the tim­ing can be more ac­cu­rate than the minute. But which min­utes does a Death Note user have to choose from? Inas­much as the Death Note is ap­par­ently in­ca­pable of in­flu­enc­ing the past or caus­ing Pratch­et­t­ian21 su­per­lu­mi­nal effects, the past is off-lim­its; but mes­sages also have to be sent in time for what­ever they are sup­posed to in­flu­ence, so one can­not afford to have a win­dow of a cen­tu­ry. If the mes­sage needs to affect some­thing within the day, then the user has a win­dow of only 60 · 24 = 1440 min­utes, which is log2(1440) = 10.49 bits; if the user has a win­dow of a year, that’s slightly bet­ter, as a death’s tim­ing down to the minute could em­body as much as log2(60 · 24 · 365) = 19 bits. (Over a decade then is 22.3 bits, etc.) If we al­low tim­ing down to the sec­ond, then a year would be 24.9 bits. In any case, it’s clear that we’re not go­ing to get more than 33 bits from the date. On the plus side, an ‘IP over Death’ pro­to­col would be su­pe­rior to —here, the worse your la­ten­cy, the more bits you could ex­tract from the pack­et’s time­stamp! on com­pres­sion schemes:

    “Yeah, but there’s more to be­ing smart than know­ing com­pres­sion schemes!” “No there’s not!” “Shoot—he knows the se­cret!!” –Ryan North
  3. the cir­cum­stances (such as the place)

    The ‘how’… has many more de­grees of free­dom. The cir­cum­stances is much more diffi­cult to cal­cu­late. We can sub­di­vide it in a lot of ways; here’s one:

    1. Lo­ca­tion (eg. lat­i­tude/­lon­gi­tude)

      Earth has ~510,072,000,000 square me­ters of sur­face area; most of it is en­tirely use­less from our per­spec­tive—if some­one is in an air­plane and dies, how on earth does one fig­ure out the ex­act square me­ter he was above? Or on the oceans? Earth has ~148,940,000,000 square me­ters of land, which is more us­able: the usual cal­cu­la­tions gives us log2(148940000000) = 37.12 bits. (Sur­prised at how sim­i­lar to the ‘who?’ bit cal­cu­la­tion this is? But 37.12 - 33 = 4.12 and 24.12 = 17.4. The SF clas­sic drew its name from the ob­ser­va­tion that the 7 bil­lion peo­ple alive in 2010 would fit in Zanz­ibar only if they stood shoul­der to shoul­der—spread them out, and mul­ti­ply that area by ~18…) This raises an is­sue that affects all 3: how much can the Death Note con­trol? Can it move vic­tims to ar­bi­trary points in, say, Siberia? Or is it lim­ited to within dri­ving dis­tance? etc. Any of those is­sues could shrink the 37 bits by a great deal.

    2. Cause Of Death

      The In­ter­na­tional Clas­si­fi­ca­tion of Dis­eases lists up­wards of 20,000 dis­eases, and we can imag­ine thou­sands of pos­si­ble ac­ci­den­tal or de­lib­er­ate deaths. But what mat­ters is what gets com­mu­ni­cat­ed: if there are 500 dis­tinct brain can­cers but the death is only re­ported as ‘brain can­cer’, the 500 count as 1 for our pur­pos­es. But we’ll be gen­er­ous and go with 20,000 for re­ported dis­eases plus ac­ci­dents, which is log2(20000) = 14.3 bits.

    3. Ac­tion Prior To Death

      Ac­tions prior to death over­laps with ac­ci­den­tal caus­es; here the se­ries does­n’t help us. Light’s early ex­per­i­ments cul­mi­nat­ing in the “L, do you know death gods love ap­ples?” seem to im­ply that ac­tions are lim­ited in en­tropy as each word took a death (as­sum­ing the or­di­nary Eng­lish vo­cab­u­lary of 50,000 words, 16 bit­s), but other plot events im­ply that hu­mans can un­der­take long com­plex plans at the or­der of Death Notes (like Mikami bring­ing the fake Death Note to the fi­nal con­fronta­tion with Near). Ac­tions be­fore death could be re­ported in great de­tail, or they could be hid­den un­der offi­cial se­crecy like the afore­men­tioned death gods men­tioned (Light uniquely priv­i­leged in learn­ing it suc­ceeded as part of L test­ing him). I can’t be­gin to guess how many dis­tinct nar­ra­tives would sur­vive trans­mis­sion or what lim­its the Note would set. We must leave this one un­de­fined: it’s al­most surely more than 10 bits, but how many?

Sum­ming, we get <33 + <19 + 17 + <37 + 14 + {?} = 120{?} bits per death.

“Bayesian Jurisprudence”

in his posthu­mous Prob­a­bil­ity The­o­ry: The Logic of Sci­ence (on ) in­cludes a chap­ter 5 on “Queer Uses For Prob­a­bil­ity The­ory”, dis­cussing such top­ics as ESP; mir­a­cles; heuris­tics & ; how vi­sual per­cep­tion is the­o­ry-laden; phi­los­o­phy of sci­ence with re­gard to New­ton­ian me­chan­ics and the famed ; horse-rac­ing & weather fore­cast­ing; and fi­nal­ly—­sec­tion 5.8, “Bayesian ju­rispru­dence”. Jay­nes’s analy­sis is some­what sim­i­lar in spirit to my above analy­sis, al­though mine is not ex­plic­itly Bayesian ex­cept per­haps in the dis­cus­sion of gen­der as elim­i­nat­ing one nec­es­sary bit.

The fol­low­ing is an ex­cerpt; see also “Bayesian Jus­tice”.

It is in­ter­est­ing to ap­ply prob­a­bil­ity the­ory in var­i­ous sit­u­a­tions in which we can’t al­ways re­duce it to num­bers very well, but still it shows au­to­mat­i­cally what kind of in­for­ma­tion would be rel­e­vant to help us do plau­si­ble rea­son­ing. Sup­pose some­one in New York City has com­mit­ted a mur­der, and you don’t know at first who it is, but you know that there are 10 mil­lion peo­ple in New York City. On the ba­sis of no knowl­edge but this, e(Guilty|X) = −70 db is the plau­si­bil­ity that any par­tic­u­lar per­son is the guilty one.

How much pos­i­tive ev­i­dence for guilt is nec­es­sary be­fore we de­cide that some man should be put away? Per­haps +40 db, al­though your re­ac­tion may be that this is not safe enough, and the num­ber ought to be high­er. If we raise this num­ber we give in­creased pro­tec­tion to the in­no­cent, but at the cost of mak­ing it more diffi­cult to con­vict the guilty; and at some point the in­ter­ests of so­ci­ety as a whole can­not be ig­nored.

For ex­am­ple, if 1000 guilty men are set free, we know from only too much ex­pe­ri­ence that 200 or 300 of them will pro­ceed im­me­di­ately to in­flict still more crimes upon so­ci­ety, and their es­cap­ing jus­tice will en­cour­age 100 more to take up crime. So it is clear that the dam­age to so­ci­ety as a whole caused by al­low­ing 1000 guilty men to go free, is far greater than that caused by falsely con­vict­ing one in­no­cent man.

If you have an emo­tional re­ac­tion against this state­ment, I ask you to think: if you were a judge, would you rather face one man whom you had con­victed false­ly; or 100 vic­tims of crimes that you could have pre­vent­ed? Set­ting the thresh­old at +40 db will mean, crude­ly, that on the av­er­age not more than one con­vic­tion in 10,000 will be in er­ror; a judge who re­quired ju­ries to fol­low this rule would prob­a­bly not make one false con­vic­tion in a work­ing life­time on the bench.

In any event, if we took +40 db start­ing out from −70 db, this means that in or­der to en­sure a con­vic­tion you would have to pro­duce about 110 db of ev­i­dence for the guilt of this par­tic­u­lar per­son. Sup­pose now we learn that this per­son had a mo­tive. What does that do to the plau­si­bil­ity for his guilt? Prob­a­bil­ity the­ory says


since , i.e. we con­sider it quite un­likely that the crime had no mo­tive at all. Thus, the [im­por­tance] of learn­ing that the per­son had a mo­tive de­pends al­most en­tirely on the prob­a­bil­ity that an in­no­cent per­son would also have a mo­tive.

This ev­i­dently agrees with our com­mon sense, if we pon­der it for a mo­ment. If the de­ceased were kind and loved by all, hardly any­one would have a mo­tive to do him in. Learn­ing that, nev­er­the­less, our sus­pect did have a mo­tive, would then be very [im­por­tant] in­for­ma­tion. If the vic­tim had been an un­sa­vory char­ac­ter, who took great de­light in all sorts of foul deeds, then a great many peo­ple would have a mo­tive, and learn­ing that our sus­pect was one of them is not so [im­por­tan­t]. The point of this is that we don’t know what to make of the in­for­ma­tion that our sus­pect had a mo­tive, un­less we also know some­thing about the char­ac­ter of the de­ceased. But how many mem­bers of ju­ries would re­al­ize that, un­less it was pointed out to them?

Sup­pose that a very en­light­ened judge, with pow­ers not given to judges un­der present law, had per­ceived this fact and, when tes­ti­mony about the mo­tive was in­tro­duced, he di­rected his as­sis­tants to de­ter­mine for the jury the num­ber of peo­ple in New York City who had a mo­tive. If this num­ber is then

and equa­tion (5-38) re­duces, for all prac­ti­cal pur­pos­es, to


You see that the pop­u­la­tion of New York has can­celed out of the equa­tion; as soon as we know the num­ber of peo­ple who had a mo­tive, then it does­n’t mat­ter any more how large the city was. Note that (5-39) con­tin­ues to say the right thing even when is only 1 or 2.

You can go on this way for a long time, and we think you will find it both en­light­en­ing and en­ter­tain­ing to do so. For ex­am­ple, we now learn that the sus­pect was seen near the scene of the crime shortly be­fore. From Bayes’ the­o­rem, the [im­por­tance] of this de­pends al­most en­tirely on how many in­no­cent per­sons were also in the vicin­i­ty. If you have ever been told not to trust Bayes’ the­o­rem, you should fol­low a few ex­am­ples like this a good deal fur­ther, and see how in­fal­li­bly it tells you what in­for­ma­tion would be rel­e­vant, what ir­rel­e­vant, in plau­si­ble rea­son­ing.22

In re­cent years there has grown up a con­sid­er­able lit­er­a­ture on Bayesian ju­rispru­dence; for a re­view with many ref­er­ences, see Vi­g­naux and Robert­son (1996) [This is ap­par­ently In­ter­pret­ing Ev­i­dence: Eval­u­at­ing Foren­sic Sci­ence in the Court­room –Ed­i­tor].

Even in sit­u­a­tions where we would be quite un­able to say that nu­mer­i­cal val­ues should be used, Bayes’ the­o­rem still re­pro­duces qual­i­ta­tively just what your com­mon sense (after per­haps some med­i­ta­tion) tells you. This is the fact that George Polya demon­strated in such o ex­haus­tive de­tail that the present writer was con­vinced that the con­nec­tion must be more than qual­i­ta­tive.

  1. In fact, every sin­gle per­son men­tioned in my Ter­ror­ism is not Effec­tive is male, and this seems to be true of the full as well.↩︎

  2. This rea­son­ing would be wrong in the case of , but Misa is an ab­surd char­ac­ter—a Gothic lolita pop star who falls in love with Light through an ex­tra­or­di­nary co­in­ci­dence and does­n’t flinch at any­thing, even sac­ri­fic­ing 75% of her lifes­pan or her mem­o­ries; hence it’s not sur­pris­ing to learn on Wikipedia from the au­thor that the mo­ti­va­tion for her char­ac­ter was to avoid a “bor­ing” al­l-male cast and be “a cute fe­male”. (Death Note is not im­mune to the Rule of Cool or Rule of Sexy!)↩︎

  3. Acausal­ity is an odd sort of new con­cept in , pri­mar­ily dis­cussed in , chap­ters 5–7, and on Less­Wrong.­com.↩︎

  4. My first so­lu­tion in­volved sex re­as­sign­ment surgery, but that makes the sit­u­a­tion worse, as trans­sex­u­als are so rare that an L in­tel­li­gent enough to an­tic­i­pate these ul­tra­-ra­tional Death Note users would in­stantly gain a huge clue: just check every­one on the surgery lists. Any­way, most Death Note users would prob­a­bly pre­fer the pass­ing-it-on so­lu­tion.↩︎

  5. This ap­plies to many other ac­tiv­i­ties like Twit­ter posts or Google search­es; eg. blog­ger mu­flax ob­served the same clear cir­ca­dian rhythms in his Google searches by hour.↩︎

  6. See the 2011 pa­per, .↩︎

  7. You can steal in­for­ma­tion through JS or CSS, and an­a­lyz­ing the his­tory for in­fer­ring de­mo­graph­ics is al­ready patented.↩︎

  8. You can try your own browser live at the ’s Panop­ticlick.↩︎

  9. Fel­ten & Schnei­der 2000, “Tim­ing At­tacks on Web Pri­vacy”↩︎

  10. See also the re­searchers’ blog.↩︎

  11. Cov­er­age of this de-anonymiza­tion al­go­rithm gen­er­ally linked it to rat­ings, but the au­thors are clear—you could have those rat­ings from any source, there’s noth­ing spe­cial about IMDb aside from it be­ing pub­lic and on­line.↩︎

  12. This sounds like some­thing that ought to be , and while the graph iso­mor­phism prob­lem is known to be in NP, it is al­most unique in be­ing like —it may be easy or hard, there is no proof ei­ther way. In prac­tice, large re­al-world graphs tend to be effi­cient to solve.↩︎

  13. From the pa­per’s ab­stract:

    [we] de­velop a new re-i­den­ti­fi­ca­tion al­go­rithm tar­get­ing anonymized so­cial-net­work graphs. To demon­strate its effec­tive­ness on re­al-world net­works, we show that a third of the users who can be ver­i­fied to have ac­counts on both Twit­ter, a pop­u­lar mi­croblog­ging ser­vice, and Flickr, an on­line pho­to-shar­ing site, can be re-i­den­ti­fied in the anony­mous Twit­ter graph with only a 12% er­ror rate. Our de-anonymiza­tion al­go­rithm is based purely on the net­work topol­o­gy, does not re­quire cre­ation of a large num­ber of dummy “sybil” nodes, is ro­bust to noise and all ex­ist­ing de­fens­es, and works even when the over­lap be­tween the tar­get net­work and the ad­ver­sary’s aux­il­iary in­for­ma­tion is small.

  14. eg. 97% of the Cam­bridge, Mass­a­chu­setts vot­ers could be iden­ti­fied with birth-date and zip code, and 29% by birth-date and just gen­der.↩︎

  15. See “Bub­ble Trou­ble: Off-Line De-Anonymiza­tion of Bub­ble Forms”, USENIX 2011S Se­cu­rity Sym­po­sium; from “New Re­search Re­sult: Bub­ble Forms Not So Anony­mous”:

    If bub­ble mark­ing pat­terns were com­pletely ran­dom, a clas­si­fier could do no bet­ter than ran­domly guess­ing a test set’s cre­ator, with an ex­pected ac­cu­racy of 1⁄92 ~ 1%. Our clas­si­fier achieves over 51% ac­cu­ra­cy. The clas­si­fier is rarely far off: the cor­rect an­swer falls in the clas­si­fier’s top three guesses 75% of the time (vs. 3% for ran­dom guess­ing) and its top ten guesses more than 92% of the time (vs. 11% for ran­dom guess­ing).

  16. See also , Cut­ler & Kulis 2018/ or , or , for ex­am­ples of what or­di­nary use of so­cial me­dia or me­dia con­sump­tion can leak.↩︎

  17. Arvind Narayanan and Vi­taly Shmatikov brusquely sum­ma­rize the im­pli­ca­tions of their de-anonymiza­tion:

    So, what’s the so­lu­tion?

    We do not be­lieve that there ex­ists a tech­ni­cal so­lu­tion to the prob­lem of anonymity in so­cial net­works. Specifi­cal­ly, we do not be­lieve that any graph trans­for­ma­tion can (a) sat­isfy a ro­bust de­fi­n­i­tion of pri­va­cy, (b) with­stand de-anonymiza­tion at­tacks de­scribed in our pa­per, and (c) pre­serve the util­ity of the graph for com­mon data-min­ing and ad­ver­tis­ing pur­pos­es. There­fore, we ad­vo­cate non-tech­ni­cal so­lu­tions.

    So, the de-anonymiz­ing just hap­pens be­hind closed doors:

    …re­searchers don’t have the in­cen­tive for deanonymiza­tion any­more. On the other hand, if ma­li­cious en­ti­ties do it, nat­u­rally they won’t talk about it in pub­lic, so there will be no PR fall­out. Reg­u­la­tors have not been very ag­gres­sive in in­ves­ti­gat­ing anonymized data re­leases in the ab­sence of a pub­lic out­cry, so that may be a neg­li­gi­ble risk. Some have ques­tioned whether deanonymiza­tion in the wild is ac­tu­ally hap­pen­ing. I think it’s a bit silly to as­sume that it is­n’t, given the eco­nomic in­cen­tives. Of course, I can’t prove this and prob­a­bly never can. No com­pany do­ing it will pub­licly talk about it, and the pri­vacy harms are so in­di­rect that ty­ing them to a spe­cific data re­lease is next to im­pos­si­ble. I can only offer anec­dotes to ex­plain my po­si­tion: I have been ap­proached mul­ti­ple times by or­ga­ni­za­tions who wanted me to deanonymize a data­base they’d ac­quired, and I’ve had friends in differ­ent in­dus­tries men­tion ca­su­ally that what they do on a daily ba­sis to com­bine differ­ent data­bases to­gether is es­sen­tially deanonymiza­tion.

    In gen­er­al, there’s no clear dis­tinc­tion be­tween ‘use­ful’ and ‘use­less’ in­for­ma­tion from the per­spec­tive of iden­ti­fy­ing/break­ing pri­va­cy/re­vers­ing anonymiza­tion (em­pha­sis added):

    ‘Qua­si­-i­den­ti­fier’ is a no­tion that arises from at­tempt­ing to see some at­trib­utes (such as ZIP code) but not oth­ers (such as tastes and be­hav­ior) as con­tribut­ing to re-i­den­ti­fi­a­bil­i­ty. How­ev­er, the ma­jor les­son from the re-i­den­ti­fi­ca­tion pa­pers of the last few years has been that any in­for­ma­tion at all about a per­son can be po­ten­tially used to aid re-i­den­ti­fi­ca­tion.

  18. But hey, at least the lack of pri­vacy is two-way and the pub­lic can male­fac­tors like the gov­ern­ment, as ar­gues is the best out­come.

    But wait, Wik­ileaks has re­vealed the mas­sive ex­pan­sion of Amer­i­can gov­ern­ment se­crecy due to the War on Ter­ror and even the sup­posed friend of trans­paren­cy, Pres­i­dent Oba­ma, has presided over an ex­pan­sion of Pres­i­dent George W. Bush’s se­crecy pro­grams and crack­downs on of all stripes? Oh. Too bad about that, I guess.↩︎

  19. Given the ex­tremely high global stakes and ap­par­ent im­pos­si­bil­ity of the mur­ders in­di­cat­ing that L is deeply ig­no­rant of ex­tremely im­por­tant in­for­ma­tion about what is go­ing on, a more prag­matic L would have sim­ply kid­napped & tor­tured or as­sas­si­nated Light as soon as L be­gan to se­ri­ously sus­pect Light.↩︎

  20. I have since seen ex­am­ples of at­tempt­ing to cor­re­late ac­tiv­ity times with lo­ca­tion on the dark­net mar­kets and else­where, such as try­ing to in­fer the time­zones of Dread Pi­rate Roberts (USA) and Satoshi Nakamoto (?).↩︎

  21. , :

    The only things known to go faster than or­di­nary light is monar­chy, ac­cord­ing to the philoso­pher Ly Tin Wee­dle. He rea­soned like this: you can’t have more than one king, and tra­di­tion de­mands that there is no gap be­tween kings, so when a king dies the suc­ces­sion must there­fore pass to the heir in­stan­ta­neously. Pre­sum­ably, he said, there must be some el­e­men­tary par­ti­cles—kingons, or pos­si­bly queon­s—that do this job, but of course suc­ces­sion some­times fails if, in mid-flight, they strike an an­ti-par­ti­cle, or re­pub­li­con. His am­bi­tious plans to use his dis­cov­ery to send mes­sages, in­volv­ing the care­ful tor­tur­ing of a small king in or­der to mod­u­late the sig­nal, were never fully ex­panded be­cause, at that point, the bar closed.

  22. “Note that in these cases we are try­ing to de­cide, from scraps of in­com­plete in­for­ma­tion, on the truth of an Aris­totelian propo­si­tion; whether the de­fen­dant did or did not com­mit some well-de­fined ac­tion. This is the sit­u­a­tion an is­sue of fact for which prob­a­bil­ity the­ory as logic is de­signed. But there are other le­gal sit­u­a­tions quite differ­ent; for ex­am­ple, in a med­ical mal­prac­tice suit it may be that all par­ties are agreed on the facts as to what the de­fen­dant ac­tu­ally did; the is­sue is whether he did or did not ex­er­cise rea­son­able judg­ment. Since there is no offi­cial, pre­cise de­fi­n­i­tion of ‘rea­son­able judg­ment’, the is­sue is not the truth of an Aris­totelian propo­si­tion (how­ev­er, if it were es­tab­lished that he will­fully vi­o­lated one of our Chap­ter 1 desider­ata of ra­tio­nal­i­ty, we think that most ju­ries would con­vict him). It has been claimed that prob­a­bil­ity the­ory is ba­si­cally in­ap­plic­a­ble to such sit­u­a­tions, and we are con­cerned with the par­tial truth of a non-Aris­totelian propo­si­tion. We sug­gest, how­ev­er, that in such cases we are not con­cerned with an is­sue of truth at all; rather, what is wanted is a value judg­ment. We shall re­turn to this topic later (Chap­ters 13, 18).”↩︎