Kodi SynthID, watermark of Artificial Intelligence ndi chiyani?

Kuwonekera kwa AI yobereka kwalimbikitsa kupanga zithunzi, zolemba, zomvetsera, ndi mavidiyo pamlingo womwe sunawonekepo, ndipo ndi izo, kukayikira za chiyambi chawo kwakula; m'nkhani ino, Dziwani ngati zinthu zidapangidwa kapena zasinthidwa ndi mtundu imakhala chinsinsi cha kudalirika kwa digito. SynthID ikhoza kukhala yankho lalikulu.

Awa ndi malingaliro a Google DeepMind, a banja la njira "zosaoneka" za watermarking zomwe zimalowetsedwa mwachindunji muzinthu zopangidwa ndi AI kuti zithandizire kutsimikizira kotsatira popanda kuwononga khalidwe lomwe anthu amawaona.

Kodi SynthID ndi chiyani ndipo cholinga chake ndi chiyani?

Google imalongosola SynthID ngati chida chothandizira watermark yeniyeni yazinthu zopangidwa ndi AI, yopangidwa kuti ilimbikitse kuwonekera komanso kutsata. Sizimangotengera mtundu umodzi: zimaphatikiza zithunzi, zomvera, zolemba, ndi makanema, kotero kuti njira imodzi yaukadaulo itha kugwiritsidwa ntchito pamitundu yosiyanasiyana yaza media.

Mu Google ecosystem imagwiritsidwa ntchito kale m'njira zingapo:

• M'mawu, mbendera ikugwira ntchito ku mayankho a Gemini.
• Mu audio, imagwiritsidwa ntchito ndi mtundu wa Lyria komanso ndi zinthu monga kupanga ma podikasiti kuchokera ku zolemba mu Notebook LM.
• En kanema, imaphatikizidwa muzolengedwa za Veo, chitsanzo chomwe chimatha kupanga zojambula mu 1080p.

Muzochitika zonse watermark Ndi imperceptible, ndipo anapangidwa kuti kupirira kusinthidwa pafupipafupi monga kuponderezana, kusintha kwa rhythm pakudula kwamawu kapena mavidiyo, popanda kuchepetsa khalidwe.

Kupitilira ukadaulo, cholinga chake chothandiza ndi chodziwikiratu: thandizani kusiyanitsa zinthu zopangidwa ndi zomwe zimapangidwa popanda AI, kuti ogwiritsa ntchito, atolankhani ndi mabungwe athe kupanga zisankho zodziwika bwino pazakudya ndi kugawa zomwe zili.

Momwe malemba a watermark (SynthID Text) amagwirira ntchito

M'malo mwake, SynthID Text imakhala ngati a logit purosesa zomwe zimakokera mupaipi yachitsanzo cha chinenero pambuyo pa zosefera zanthawi zonse (Top-K ndi Top-P). Purosesa iyi imasinthira mobisa mawerengero amitundu ndi a pseudorandom ntchito g, kuyika zambiri m'njira yotheka popanda kuyika zinthu zakale zowoneka mumayendedwe kapena mtundu wa mawuwo.

Chotsatira chake ndi mawu omwe, poyang'ana koyamba, amasunga khalidwe, mwatsatanetsatane ndi fluidity, koma yomwe imaphatikizapo ndondomeko ya ziwerengero zomwe zimazindikirika ndi zotsimikizira zophunzitsidwa.

Kupanga malemba ndi watermark sikofunikira phunzitsanso chitsanzo: ingoperekani kasinthidwe kwa njira .generate() ndi yambitsa SynthID Text's logit processor. Izi zimathandizira kutengera ana komanso kulola kuyesa ndi zitsanzo zomwe zatumizidwa kale.

Zokonda za watermark zili ndi magawo awiri ofunikira: keys y ngram_len. Makiyi ndi mndandanda wazinthu zapadera, zosawerengeka zomwe zimagwiritsidwa ntchito polemba mawu pogwiritsa ntchito g ntchito; kutalika kwa mndandandawo kumawongolera kuchuluka kwa "zigawo" za watermarking zomwe zimagwiritsidwa ntchito. Pakadali pano, ngram_len Imayika malire pakati pa kuzindikirika ndi kulimba pakusintha: mayendedwe apamwamba amapangitsa kuzindikira kukhala kosavuta koma kumapangitsa chisindikizo kukhala pachiwopsezo chosintha; mtengo wa 5 umagwira ntchito ngati poyambira.

Kuphatikiza apo, SynthID Text imagwiritsa ntchito a sampuli tebulo ndi katundu awiri: sampling_table_size y sampling_table_seed. Kukula kwa osachepera 2 ^ 16 kumalimbikitsidwa kuwonetsetsa kuti ntchito g ikuchita mokhazikika komanso mosakondera poyesa, poganizira kuti kukula kwakukulu kumatanthauza kukumbukira zambiri panthawi yoyembekezera. Mbewuyo imatha kukhala nambala iliyonse, zomwe zimathandizira kuberekana m'malo owunika.

Pali nuance yofunika kuwongolera chizindikiro: mobwerezabwereza n-gram m'mbiri yaposachedwa ya nkhaniyo (yotanthauzidwa ndi context_history_size) sizinalembedwe, zomwe zimakonda kudziwika kwa chizindikirocho m'malemba ena onse ndikuchepetsa zolakwika zomwe zimagwirizanitsidwa ndi kubwereza kwachibadwa kwa chinenerocho.

Pachitetezo, kasinthidwe ka watermark kalikonse (kuphatikiza makiyi ake, mbewu ndi magawo) ziyenera kusungidwa mwachinsinsiNgati makiyi awa atayikira, anthu ena akhoza kutengera mtunduwo mosavuta kapena, choyipa kwambiri, kuyesa kuwusintha ndi chidziwitso chonse cha kapangidwe kake.

Momwe mungadziwire: kutsimikizira kwapang'onopang'ono kokhala ndi malire

Kutsimikizira kwa watermark m'mawu sikunaphatikizidwe, koma chothekaGoogle imasindikiza chowunikira cha Bayesian pa Transformers ndi GitHub kuti, pambuyo posanthula ndondomeko ya zolembazo, imabweza zigawo zitatu zomwe zingatheke: ndi chizindikiro, palibe mtundu o wosatsimikizaKutulutsa kwamtundu wa ternary kumapangitsa kuti ntchitoyi isinthidwe kuzinthu zosiyanasiyana zowopsa komanso zololera zolakwika.

Khalidwe la wotsimikizira limasinthidwa ndi malire awiri zomwe zimawongolera kuchuluka kwa zinthu zabodza komanso zoyipa zabodza. Mwa kuyankhula kwina, mutha kuwerengera momwe mukufuna kuti chidziwitsocho chikhale chokhwima, kusiya kukhudzika kwa kulondola kapena mosemphanitsa kutengera momwe mumagwiritsira ntchito, chinthu chofunikira kwambiri zowongolera, kuwongolera kapena kafukufuku wamkati.

Ngati zitsanzo zingapo zimagawana zomwezo chizindikiro, mutha kugawana nawo kasinthidwe ka mtundu womwewo ndi chowunikira chomwecho, bola ngati maphunziro a verifier akuphatikizapo zitsanzo za onse. Izi zimapangitsa kukhala kosavuta kupanga "ma watermark wamba" m'mabungwe omwe ali ndi ma LLM angapo.

Chowunikiracho chikaphunzitsidwa, mabungwe amatha kusankha momwe angawonekere: sungani kwathunthu payekha, perekani mwanjira ina zachinsinsi kudzera mu API, kapena kuimasula m'njira anthu kutsitsa ndikugwiritsa ntchito ndi anthu ena. Chisankhocho chimadalira mphamvu ya bungwe lirilonse, kuopsa kwa kayendetsedwe ka kayendetsedwe kake, ndi njira zowonekera.

Watermark pazithunzi, zomvera ndi makanema

Chizindikiro ichi chapangidwa kuti chikhale chokhalitsa wamba masinthidwe monga kudulira, kusintha kukula, kuzungulira, kusintha mtundu, kapena zithunzi zowonera, popanda kufunikira kosunga metadata. Poyamba, kugwiritsidwa ntchito kwake kunaperekedwa kudzera Chithunzi mu Vertex AI, pomwe ogwiritsa ntchito angasankhe kuyatsa watermark popanga zomwe zili.

Mu audio, mtundu ndi osamveka ndipo imathandizira magwiridwe antchito wamba monga kupsinjika kwa MP3, kuwonjezera phokoso, kapena kusintha liwiro losewera. Google ikuphatikiza ndi Lyria ndi mu Notebook LM-based features, kulimbikitsa chizindikiro ngakhale pamene fayilo ikudutsa mitsinje yofalitsa yotayika.

Muvidiyoyi, njirayo imafanizira njira yachithunzi: chizindikirocho chikuphatikizidwa mu ma pixel a chimango chilichonse, mosazindikira, ndipo imakhala yokhazikika motsutsana ndi zosefera, kusintha kwa kuchuluka kwa zotsitsimutsa, kukanikiza kapena recortes. Makanema opangidwa ndi Ndikuwona Zida ngati VideoFX zimaphatikizira chizindikirochi pakulenga, kuchepetsa chiopsezo chochotsa mwangozi pazosintha zotsatila.

Sampling ma algorithms ndi kulimba kwa chisindikizo chalemba

Mtima wa SynthID Text ndi wake sampuli algorithm, yomwe imagwiritsa ntchito kiyi (kapena seti ya makiyi) kuti ipereke zigoli zachisawawa pa chizindikiro chilichonse. Otsatira amatengedwa kuchokera ku kagawidwe kachitsanzo (pambuyo pa Top-K/Top-P) ndikuyikidwa mu "mpikisano" wotsatira maulendo ochotsa, mpaka chizindikiro chopambana kwambiri chisankhidwe molingana ndi ntchito g.

Njira yosankhidwa iyi imakomera ndondomeko yomaliza yowerengera za kuthekera kukhala ndi chizindikiro cha mtundu, koma popanda kukakamiza zosankha zachilendo. Malinga ndi maphunziro ofalitsidwa, njirayi imapangitsa kuti zikhale zovuta kufufuta, kunamiza, kapena kutembenuza chisindikizo, nthawi zonse mkati mwa malire oyenera motsutsana ndi otsutsa ndi nthawi ndi chilimbikitso.

Kukhazikitsa bwino ndi machitidwe achitetezo

• Ngati mukugwiritsa ntchito SynthID Text, sankhani kasinthidwe ngati chinsinsi chopanga: Sungani makiyi ndi mbewu mu manejala wotetezedwa, tsatirani ziwongolero zolowera, ndipo lolani kuti azisinthasintha pafupipafupi. Kupewa kutayikira kumachepetsa kuukira komwe kumayesa mosinthana ndi mainjiniya.
• Pangani dongosolo kuyang'anira pa chojambulira chanu: sungani mitengo yabodza / yoyipa, sinthani malire malinga ndi zomwe zikuchitika ndikusankha zomwe mukufuna kudziwa. kufotokoza (zachinsinsi, zachinsinsi kudzera pa API, kapena zapagulu) zokhala ndi malamulo omveka bwino komanso machitidwe. Ndipo ngati mitundu ingapo imagawana chizindikiro, lingalirani maphunziro a wamba chodziwira ndi zitsanzo za zonsezi kuti zikhale zosavuta kukonza.
• Pamlingo wa magwiridwe antchito, imawunika momwe zimakhudzira sampling_table_size mu kukumbukira ndi latency, ndi kusankha a ngram_len zomwe zimalinganiza kulolerana kwanu kwa zosintha ndi kufunikira kwa kuzindikira kodalirika. Kumbukirani kusaphatikiza ma n-grams (kudzera context_history_size) kuwongolera chizindikiro pamawu oyenda.

SynthID si chipolopolo chasiliva chotsutsana ndi zabodza, koma imapereka chomangira chofunikira pakumanganso mayendedwe odalirika munthawi ya AI yopangira. Mwa kuyika zizindikiro zoyambira m'mawu, zithunzi, zomvera, ndi makanema, ndikutsegulira gawo lazolemba kwa anthu ammudzi, Google DeepMind ikukankhira mtsogolo momwe zowona zitha kuwerengedwa m'njira yothandiza, yopimidwa, komanso, koposa zonse, yogwirizana ndi luso komanso mtundu wa zomwe zili.