SynthID yog dab tsi, watermark ntawm kev txawj ntse?

Qhov tshwm sim ntawm AI generative tau txhawb nqa cov duab, ntawv, audios, thiab cov yeeb yaj kiab ntawm qhov ntsuas uas tsis tau pom dua ua ntej, thiab nrog nws, kev tsis ntseeg txog lawv cov keeb kwm tau loj hlob; nyob rau hauv cov ntsiab lus no, Txheeb xyuas seb cov ntsiab lus tau tsim lossis hloov pauv los ntawm tus qauv dhau los ua tus yuam sij rau kev ntseeg digital. SynthID tuaj yeem yog qhov kev daws teeb meem zoo.

Nov yog Google DeepMind lub tswv yim, a tsev neeg ntawm "invisible" txheej txheem watermarking uas tau kos ncaj qha rau hauv AI-tsim cov ntsiab lus los pab txhawb kev txheeb xyuas tom ntej yam tsis muaj kev cuam tshuam qhov zoo ntawm tib neeg pom.

SynthID yog dab tsi thiab nws npaj rau dab tsi?

Google piav qhia SynthID ua ib qho cuab yeej rau tshwj xeeb watermark rau AI-tsim cov ntsiab lus, tsim los txhawb pob tshab thiab traceability. Nws tsis txwv rau ib hom ntawv: nws encompasses dluab, suab, ntawv nyeem, thiab video, thiaj li hais tias ib tug technical mus kom ze yuav siv tau rau ntau hom xov xwm.

Hauv Google ecosystem nws twb tau siv ntau txoj hauv kev:

• Hauv ntawv, tus chij siv rau Gemini cov lus teb.
• Hauv suab, yog siv nrog Lyria qauv thiab nrog cov yam ntxwv xws li tsim cov podcasts los ntawm cov ntawv nyeem hauv Notebook LM.
• En video, yog kev koom ua ke rau hauv Veo tsim, tus qauv muaj peev xwm tsim cov clips hauv 1080p.

Hauv txhua qhov xwm txheej cov cim dej Nws yog imperceptible, thiab tau tsim los rau tiv taus kev hloov kho tsis tu ncua xws li compression, atherosclerosis hloov hauv lub suab lossis video txiav, tsis txo qhov zoo.

Dhau li ntawm thev naus laus zis, nws lub hom phiaj siv tau yog qhov tseeb: pab kom paub qhov txawv ntawm cov khoom siv hluavtaws los ntawm cov khoom uas tsis muaj AI, yog li ntawd cov neeg siv, xov xwm thiab cov koom haum tuaj yeem txiav txim siab paub txog kev siv thiab kev faib cov ntsiab lus.

Yuav ua li cas cov ntawv watermark (SynthID Text) ua haujlwm

Hauv kev xyaum, SynthID Text ua raws li a logit processor uas hooks rau hauv cov lus qauv tsim cov raj xa dej tom qab cov lim dej ib txwm siv (Top-K thiab Top-P). Cov processor no hloov pauv hloov cov qauv qhab nia nrog a pseudorandom ua g, encoding cov ntaub ntawv nyob rau hauv tus qauv ntawm qhov yuav tshwm sim yam tsis tau qhia pom artifacts rau hauv cov style lossis zoo ntawm cov ntawv nyeem.

Qhov tshwm sim yog cov ntawv nyeem uas, thaum xub thawj siab ib muag, tswj zoo, precision thiab fluidity, tab sis uas suav nrog cov qauv kev txheeb xyuas pom tau nrog tus kws qhia paub tseeb.

Txhawm rau tsim cov ntawv nrog watermark nws tsis tsim nyog rov cob qhia tus qauv: tsuas yog muab kev teeb tsa rau txoj kev .generate() thiab qhib SynthID Text's logit processor. Qhov no yooj yim rau kev saws me nyuam thiab tso cai rau kev sim nrog cov qauv uas twb tau siv lawm.

Cov kev teeb tsa watermark suav nrog ob qhov tseem ceeb: keys y ngram_len. Cov yuam sij yog ib daim ntawv teev npe tshwj xeeb, random integers siv los ntsuas cov lus siv g muaj nuj nqi; qhov ntev ntawm daim ntawv teev npe tswj tau pes tsawg "txheej" ntawm watermarking yog thov. Lub caij no, ngram_len Teem qhov sib npaug ntawm kev kuaj pom thiab muaj zog rau kev hloov pauv: cov txiaj ntsig siab dua ua kom pom tau yooj yim dua tab sis ua rau lub foob muaj kev cuam tshuam rau kev hloov pauv; tus nqi ntawm 5 ua haujlwm zoo li qhov pib.

Tsis tas li ntawd, SynthID Text siv a Sampling rooj nrog ob lub zog: sampling_table_size y sampling_table_seed. Qhov loj me me ntawm tsawg kawg yog 2^16 raug pom zoo los xyuas kom meej tias qhov kev ua haujlwm g coj tus cwj pwm kom ruaj khov thiab tsis ncaj ncees thaum ua piv txwv, suav nrog qhov ntawd. qhov loj dua txhais tau tias nco ntau dua thaum inference. Cov noob tuaj yeem yog txhua tus lej, uas pab txhawb kev rov ua dua tshiab hauv qhov chaw ntsuas.

Muaj ib qho tseem ceeb nuance los txhim kho lub teeb liab: rov n-grams nyob rau hauv keeb kwm tsis ntev los no ntawm cov ntsiab lus (txhais los ntawm context_history_size) tsis muaj cim, uas nyiam qhov kev tshawb pom ntawm cov cim nyob rau hauv cov ntawv nyeem thiab txo cov lus tsis tseeb uas txuas rau cov lus rov qab ntawm cov lus.

Rau kev ruaj ntseg, txhua qhov kev teeb tsa watermark (nrog rau nws cov yawm sij, noob thiab tsis muaj) yuav tsum muab khaws cia tsegYog tias cov yuam sij no tau xau, cov neeg thib peb tuaj yeem yooj yim rov ua dua lub npe lossis, tsis zoo dua, sim ua kom nws nrog kev paub tag nrho ntawm nws cov qauv.

Yuav kuaj li cas: kev kuaj xyuas qhov tseeb nrog qhov pib

Kev pov thawj ntawm cov cim dej hauv cov ntawv tsis yog binary, tab sis probabilisticGoogle tshaj tawm Bayesian detector ntawm Transformers thiab GitHub uas, tom qab tshuaj xyuas tus qauv ntawm cov ntawv nyeem, rov qab peb lub xeev: hom lag luam, tsis muaj npe o tsis paub meejQhov kev tso tawm ternary no tso cai rau kev ua haujlwm kom raug hloov kho rau cov kev pheej hmoo sib txawv thiab kev ua yuam kev kam rau ua.

Tus cwj pwm ntawm tus neeg txheeb xyuas yog configurable los ntawm ob qhov chaw uas tswj tus nqi ntawm qhov tsis zoo thiab qhov tsis zoo tsis tseeb. Hauv lwm lo lus, koj tuaj yeem ntsuas qhov nruj npaum li cas koj xav kom qhov kev tshawb pom, ua rau muaj kev cuam tshuam rau qhov raug lossis hloov pauv nyob ntawm koj cov ntaub ntawv siv, qee yam tshwj xeeb tshaj yog siv hauv editorial ib puag ncig, moderation los yog kev soj ntsuam sab hauv.

Yog tias ob peb tus qauv sib koom ua ke tokenizer, kuj tuaj yeem faib cov tib hom configuration thiab tib lub ntes, tsuav yog tus neeg txheeb xyuas qhov kev cob qhia muaj xws li cov piv txwv ntawm txhua tus ntawm lawv. Qhov no ua rau nws yooj yim dua los tsim "hom watermarks" hauv cov koom haum nrog ntau LLMs.

Thaum lub tshuab ntes tau raug cob qhia, cov koom haum tuaj yeem txiav txim siab nws qib ntawm qhov raug: khaws cia ntiag tug kiag li, muab nws nyob rau hauv ib txoj kev ib nrab ntiav los ntawm API, los yog tso tawm rau hauv ib txoj kev pej xeem rau download thiab siv los ntawm peb tog. Qhov kev xaiv yog nyob ntawm txhua qhov chaw ua haujlwm qhov peev txheej ua haujlwm, kev tswj hwm kev pheej hmoo, thiab cov tswv yim pom tseeb.

Watermark ntawm cov duab, suab thiab video

Hom no yog tsim los kav ntev kev hloov pauv xws li cropping, resizing, rotating, hloov xim, lossis txawm screenshots, tsis tas yuav khaws cov metadata. Thaum xub thawj, nws siv tau los ntawm Duab hauv Vertex AI, qhov twg cov neeg siv tuaj yeem xaiv qhib lub watermark thaum tsim cov ntsiab lus.

Hauv suab, lub npe yog tsis hnov ​​lus thiab txhawb kev ua haujlwm xws li MP3 compression, ntxiv suab nrov, lossis hloov kho qhov nrawm nrawm. Google integrates nws rau hauv Lyria thiab hauv Notebook LM-raws li nta, boosting lub teeb liab txawm tias thaum cov ntaub ntawv hla lossy publishing kwj.

Hauv video, txoj hauv kev rov ua dua cov duab mus kom ze: lub hom phiaj yog kos rau hauv pixels ntawm txhua tus ncej, imperceptibly, thiab tseem ruaj khov tiv thaiv cov ntxaij lim dej, kev hloov pauv hauv tus nqi refresh, compression lossis txiav. Videos generated los ntawm Kuv pom Cov cuab yeej zoo li VideoFX suav nrog cov cim no thaum lub sijhawm tsim, txo qhov kev pheej hmoo ntawm kev raug rho tawm hauv kev hloov kho tom ntej.

Sampling algorithms thiab robustness ntawm cov ntawv foob

Lub plawv ntawm SynthID Text yog nws kev sampling algorithm, uas siv tus yuam sij (los yog cov yuam sij) los muab cov qhab nia pseudo-random rau txhua qhov muaj peev xwm token. Cov neeg sib tw tau raug rho tawm los ntawm tus qauv kev faib tawm (tom qab Top-K / Top-P) thiab muab tso rau hauv "kev sib tw" tom qab tshem tawm cov kev sib tw, kom txog rau thaum cov qhab nia siab tshaj plaws tau xaiv raws li txoj haujlwm g.

Txoj kev xaiv no nyiam cov tus qauv kev txheeb cais zaum kawg ntawm qhov tshwm sim dais lub cim ntawm lub hom, tab sis tsis muaj kev yuam kev unnatural xaiv. Raws li kev tshawb fawb luam tawm, cov txheej txheem ua rau nws nyuaj lwv, dag, lossis thim rov qab lub foob, ib txwm nyob rau hauv qhov tsim nyog txwv tiv thaiv cov neeg tawm tsam nrog lub sijhawm thiab kev txhawb siab.

Kev ua tau zoo thiab kev ua haujlwm ruaj ntseg

• Yog tias koj tab tom siv SynthID Text, kho qhov teeb tsa raws li ntau lawm secret: Khaws cov yuam sij thiab cov noob rau hauv tus neeg saib xyuas kev ruaj ntseg, tswj xyuas kev nkag mus, thiab tso cai rau kev sib hloov raws sij hawm. Tiv thaiv kom tsis txhob cuam tshuam qhov chaw tawm tsam tawm tsam rov qab engineering sim.
• Tsim ib txoj kev npaj rau kev saib xyuas rau koj tus neeg ntes: sau cov nqi tsis zoo / tsis zoo, kho qhov pib raws li cov ntsiab lus thiab txiav txim siab koj txoj cai tshawb nrhiav nphav (piv txwv, semi-private ntawm API, lossis pej xeem) nrog cov qauv kev cai lij choj thiab kev ua haujlwm. Thiab yog tias muaj ntau tus qauv qhia ib qho tokenizer, xav txog kev cob qhia a common detector nrog rau cov piv txwv ntawm tag nrho cov ntawm lawv kom yooj yim txij nkawm.
• Nyob rau theem kev ua tau zoo, nws ntsuas qhov cuam tshuam ntawm sampling_table_size hauv kev nco thiab latency, thiab xaiv a ngram_len uas sib npaug koj lub siab ntev rau kev hloov kho nrog rau qhov xav tau kom ntseeg tau qhov tseeb. Nco ntsoov tshem tawm cov n-grams rov qab (ntawm context_history_size) txhawm rau txhim kho lub teeb liab hauv cov ntawv ntws.

SynthID tsis yog lub mos txwv nyiaj tiv thaiv cov ntaub ntawv tsis raug, tab sis muab lub tsev thaiv tseem ceeb rau kev rov tsim kho cov saw hlau ntawm kev ntseeg siab nyob rau hauv lub sijhawm ntawm generative AI. Los ntawm embedding provenance signals nyob rau hauv cov ntawv nyeem, duab, suab, thiab video, thiab qhib lub ntsiab lus rau lub zej zog, Google DeepMind yog thawb mus rau lub neej yav tom ntej uas muaj tseeb yuav raug soj ntsuam nyob rau hauv ib tug ua tau zoo, ntsuas, thiab, qhov tseem ceeb tshaj plaws, sib xws nrog lub tswv yim thiab zoo ntawm cov ntsiab lus.