PassGPT: Tsarin Kalmar Sirri da Samar da Kalmar Sirri tare da Manyan Hanyoyin Harshe

1. Gabatarwa

Kalmar sirri ta ci gaba da zama babbar hanyar tabbatar da ainihi saboda sauƙinta da yadda ake amfani da ita. Duk da haka, fallasa kalmar sirri yana haifar da babbar barazana, yana ba da damar kai hari da kuma bincike kan tsarin ƙirƙirar kalmar sirri na ɗan adam. Wannan takarda tana binciken aikace-aikacen Manyan Hanyoyin Harshe (LLMs) akan tsarin kalmar sirri, tare da gabatar da PassGPT. PassGPT wani LLM ne da aka horar da shi akan fallasa kalmar sirri don samarwa da kimanta ƙarfi, yana nuna babban aiki fiye da hanyoyin da suka gabata na Tsarin Adawa na Halitta (GAN) kuma yana gabatar da sabbin iyawa kamar samarwa mai jagora.

2. Hanyoyi & Tsarin Gine-gine

PassGPT an gina shi akan tsarin gine-ginen GPT-2, wanda aka daidaita don samar da kalmar sirri a matakin jeri, a matakin haruffa. Wannan hanyar ta bambanta sosai da GANs waɗanda ke samar da kalmar sirri a matsayin raka'a guda ɗaya, atomatik.

2.1. Tsarin Gine-ginen PassGPT

Tsarin ya dogara ne akan tsarin gine-ginen Transformer decoder. Yana sarrafa kalmar sirri a matsayin jerin haruffa (ko alamomi), yana koyon yuwuwar yanayin harafi na gaba idan aka yi la'akari da mahallin da ya gabata: $P(x_t | x_{PassVQT, ya haɗa da fasahohin ƙididdiga na vector don ƙara rudani (da yuwuwar bambancin) kalmar sirri da aka samar.

2.2. Samar da Kalmar Sirri Mai Jagora

Wani sabon abu mai mahimmanci shine samar da kalmar sirri mai jagora. Ta hanyar sarrafa tsarin samfurin (misali, ta amfani da yuwuwar yanayi ko ƙayyadaddun bayani), PassGPT na iya samar da kalmar sirri waɗanda suka gamsar da ƙayyadaddun abubuwan da mai amfani ya ayyana (misali, "dole ne ya ƙunshi lamba da babban harafi"), wani aiki da ba zai yiwu ba tare da GANs na yau da kullun.

2.3. Horarwa & Bayanai

An horar da tsarin akan manyan fallasa kalmar sirri ta hanyar da ba ta da kulawa, wanda ya dace da tsarin barazanar zato kalmar sirri na waje da aka saba da shi a cikin binciken tsaro.

3. Sakamakon Gwaji & Bincike

3.1. Aikin Zato Kalmar Sirri

PassGPT ya fi na baya-bayan nan manyan tsarin samarwa mai zurfi (misali, GANs) girma. Yana zato kalmar sirri da ba a taɓa gani ba kashi 20% kuma yana nuna ƙarfin haɗawa zuwa sabbin bayanan kalmar sirri da ba a gani ba yayin horo.

3.2. Rarraba Yuwuwar & Binciken Entropy

Ba kamar GANs ba, PassGPT yana ba da bayyanannen rarraba yuwuwar akan duk sararin kalmar sirri. Bincike ya nuna cewa PassGPT yana ba da ƙananan yuwuwar (babban abin mamaki) ga kalmar sirri da aka ɗauka "mai ƙarfi" ta hanyar ƙwararrun masu kimanta ƙarfi (kamar zxcvbn), yana nuna daidaito. Hakanan yana gano kalmar sirri da masu kimanta suka ɗauka mai ƙarfi amma suna da yuwuwar yanayi a ƙarƙashin tsarin, yana bayyana raunin da ke iya faruwa.

3.3. Kwatantawa da Hanyoyin da suka dogara da GAN

Samarwar jeri na PassGPT yana ba da fa'idodi fiye da GANs: 1) Bayyanannen rarraba yuwuwar, 2) Iyawar samarwa mai jagora, 3) Mafi kyawun aiki akan bayanan da ba a gani ba. Takardar ta sanya wannan a matsayin sauyin tsari daga samarwar fitarwa guda ɗaya zuwa tsarin jeri mai sarrafawa, mai yuwuwar yanayi don kalmar sirri.

4. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Ginshiƙin PassGPT shine manufar tsarin harshe mai cin gashin kansa, yana haɓaka yuwuwar bayanan horo:

$L(\theta) = \sum_{i=1}^{N} \sum_{t=1}^{T_i} \log P(x_t^{(i)} | x_{

inda $N$ shine adadin kalmar sirri, $T_i$ shine tsawon kalmar sirri $i$, $x_t^{(i)}$ shine harafi na $t$, kuma $\theta$ su ne sigogin tsarin. Samfurin don samarwa yana amfani da hanyoyi kamar samfurin saman-k ko tsakiya don daidaita bambancin da inganci. Yuwuwar cikakkiyar kalmar sirri $S$ ita ce: $P(S) = \prod_{t=1}^{|S|} P(x_t | x_{

5. Fahimtar Asali & Ra'ayin Mai Bincike

Fahimtar Asali: Haɓakar gaske na takardar ba kawai mai fashewar kalmar sirri mafi kyau ba ne; shine tsarawa na ƙirƙirar kalmar sirri a matsayin matsalar samarwar jeri mai sarrafawa. Ta hanyar amfani da hasashen alama na gaba—ma'aikacin aikin zamani na NLP—zuwa kalmar sirri, PassGPT ya wuce samarwar akwatin baƙar fata, harbi ɗaya na GANs (kamar waɗanda ke cikin CycleGAN salon fassarar hoto) zuwa cikin tsari mai bayyana, mai iya tuƙi. Wannan yana sake tsara tsaro daga kimanta ƙarfi kawai zuwa tsarin ɗan adam tsari a bayan zaɓin kalmar sirri.

Kwararar Ma'ana: Hujjar tana da ban sha'awa: 1) LLMs suna ƙware wajen ɗaukar rikice-rikice, rarraba na duniya na gaske (rubutu). 2) Kalmar sirri ƙayyadaddun yare ne, wanda ɗan adam ya samar. 3) Don haka, LLMs yakamata su yi tsarin su yadda ya kamata—wanda suke yi, suna doke GANs. 4) Yanayin jeri na LLMs yana buɗe samarwa mai jagora, babban aikace-aikace don fashewa mai sanin manufa ko gwajin ƙarfi mai himma. 5) Fitowar yuwuwar bayyananna tana ba da ma'auni kai tsaye, mai fassara don tsaro, yana haɗa tazara tsakanin hare-haren samarwa da masu kimanta ƙarfin yuwuwar yanayi.

Ƙarfi & Kurakurai: Ƙarfin ba shakka ne: babban aiki da sabon aiki. Nunin samarwa mai jagora wani babban fasaha ne, yana nuna amfanin aiki nan take. Duk da haka, binciken yana da kuskure mai mahimmanci da aka saba da shi a cikin takardun ML-don-tsaro: yana rawa a kusa da yanayin amfani biyu. Yayin da ake ambaton "haɓaka masu kimanta ƙarfi," babban amfani da aka nuna shine na kai hari (zato). Tsarin ɗa'a yana da sirara. Bugu da ƙari, yayin da ya fi GANs girma, kwatantawa da manyan kayan aikin fashewa masu ƙa'ida kamar Hashcat tare da ƙa'idodi masu ci gaba ba a bayyana ba sosai. Aikin tsarin har yanzu yana iyakance ta bayanan horonsa—fallasa—waɗanda ƙila ba su wakilci duk halayen kalmar sirri na ɗan adam ba.

Fahimta Mai Aiki: Ga masu tsaron gida, wannan ba alamar halaka ba ce amma kira ga makamai. Na farko, masu kimanta ƙarfin kalmar sirri dole su haɗa irin waɗannan yuwuwar samarwa, kamar yadda aka ba da shawarar. Kayan aiki kamar zxcvbn yakamata a sake gyara su don duba kalmar sirri daidai da yuwuwar tsarin kamar PassGPT, ba kawai ƙa'idodin tsaye ba. Na biyu, ƙungiyoyin ja yakamata su karɓi wannan hanyar nan take don binciken cikin gida; samarwa mai jagora ya dace don gwada yarda da takamaiman manufofin kalmar sirri. Na uku, wannan binciken ya tabbatar da buƙatar motsawa bayan kalmar sirri. Idan LLM zai iya yin tsarin su da kyau haka, entropy na dogon lokaci yana rushewa. Zuba jari a cikin FIDO2/WebAuthn da maɓallan shiga ya zama mafi gaggawa. Abin da za a ɗauka: Kula da PassGPT ba a matsayin mai fashewa ba, amma a matsayin mafi daidaitaccen na'urar kwaikwayon raunin kalmar sirri na ɗan adam da aka gina tukuna. Yi amfani da shi don gyara tsaron ku kafin abokin gaba ya yi.

6. Tsarin Bincike: Misalin Hali

Yanayi: Manufar kamfani tana buƙatar kalmar sirri tare da aƙalla babban harafi ɗaya, lamba ɗaya, da harafi na musamman ɗaya. Mai fashewar kalmar sirri na al'ada mai ƙa'ida zai iya amfani da ƙa'idodin ɓarna. GAN zai yi wahala don samar da kalmar sirri masu bin ka'ida kawai.

Hanyar Samar da PassGPT Mai Jagora:

1. Ƙayyadaddun Ƙuntatawa: Ayyana abin rufe fuska ko dabaru don tsarin samfurin don tilasta matsayi na nau'in harafi.
2. Samfurin da aka Ƙuntata: Yayin samar da kowane harafi $x_t$ na cin gashin kansa, ana tace rarraba samfurin ko a nuna son kai don ba da damar haruffa daga saitin da ya gamsar da sauran buƙatun manufofin (misali, idan ba a samar da lamba ta matsayi $t$ ba, ƙara yawan yuwuwar akan lambobi).
3. Fitarwa: Tsarin yana samar da jerin kamar "C@t9Lover" ko "F1r3Tr#ck" waɗanda duka suna da yuwuwar yanayi (wanda aka koya daga fallasa) kuma suna bin manufofin.

Wannan yana nuna yadda za a iya amfani da PassGPT don gwajin tsaro mai sanin manufa, yana samar da mafi yuwuwar raunin kalmar sirri waɗanda har yanzu suna wucewa gwajin manufa, gano ɓangarorin manufa.

7. Duban Aikace-aikace & Hanyoyin Gaba

Gajeren lokaci (shekaru 1-2):

• Haɗa tsarin kamar PassGPT cikin masu kimanta ƙarfin kalmar sirri na kasuwanci da buɗe tushe (misali, a matsayin kayan haɗi don zxcvbn).
• Karɓa ta tsarin gwajin shiga don mafi inganci, hankali na fesa kalmar sirri da kwaikwayon ƙarfi.
• Haɓaka masu samar da "kalmar sirri na adawa" don gwada tsarin ta hanyar ƙirƙirar kalmar sirri waɗanda suke da ƙarfi ta hanyar ma'auni na al'ada amma suna da yuwuwar ƙarƙashin tsarin LLM.

Matsakaicin lokaci (shekaru 3-5):

• Tsarin gauraye da ke haɗa yuwuwar LLM tare da ma'aunin entropy na sirri na al'ada don maki ƙarfi mai fuskoki da yawa.
• Aikace-aikacen binciken kalmar sirri na ainihi, wanda ya dogara da API don kamfanoni, kimanta sabbin kalmar sirri daidai da ci gaba da sabunta tsarin samarwa da aka horar da shi akan sabbin fallasa.
• Aikace-aikace bayan kalmar sirri mai bayyananne: tsarin alamu a cikin jimlolin sirri, amsoshin tambayoyin tsaro, har ma da jerin halayen mai amfani don gano abin da ba na al'ada ba.

Dogon lokaci & iyakokin bincike:

• Bincika tsarin watsawa ko wasu gine-ginen zamani na gaba don tsarin kalmar sirri.
• Hanyoyin koyon tarayya don horar da tsarin kalmar sirri akan bayanan keɓaɓɓu, masu zaman kansu don inganta ɗaukar hoto ba tare da tsakiyar fallasa masu mahimmanci ba.
• Yin amfani da waɗannan tsare-tsare don ƙirƙirar sabbin tsare-tsaren ƙirƙirar kalmar sirri, masu jure wa LLM, yana iya yin amfani da tsarin samarwa don "alurar" masu amfani daga tsarin da ake iya hasashen.

Hanyar ƙarshe, kamar yadda aka nuna ta nasarar takardar, shine maye gurbin ƙa'idodin kalmar sirri na dabara da hankali tare da tsarin tsaro na bayanai, mai yuwuwar yanayi.

8. Nassoshi

1. Rando, J., Perez-Cruz, F., & Hitaj, B. (2023). PassGPT: Tsarin Kalmar Sirri da (Jagorancin) Samarwa tare da Manyan Hanyoyin Harshe. arXiv preprint arXiv:2306.01545v2.
2. Goodfellow, I., et al. (2014). Tsarin Adawa na Halitta. Ci gaba a cikin Tsarin Bayanai na Jijiyoyi.
3. Zhu, J.-Y., et al. (2017). Fassarar Hoto-zuwa-Hoto mara Haɗin gwiwa ta amfani da Tsarin Adawa Mai Daidaituwa na Zagaye. IEEE Taron Ƙwararrun Kwamfuta na Duniya (ICCV).
4. Vaswani, A., et al. (2012017). Hankali Duk Abinda Kake Bukata. Ci gaba a cikin Tsarin Bayanai na Jijiyoyi.
5. Melicher, W., et al. (2016). Sauri, Sirara, da Daidaito: Yin Tsarin Yiwuwar Kalmar Sirri ta Amfani da Hanyoyin Sadarwa na Jijiyoyi. Taron Tsaro na USENIX.
6. Weir, M., et al. (2009). Fashewar Kalmar Sirri ta Amfani da Nahawu na Mahallin Kyauta mai Yuwuwar Yanayi. IEEE Taron kan Tsaro da Keɓantawa.
7. Ƙungiyar FIDO. (2023). Ƙayyadaddun FIDO2/WebAuthn. An samo daga https://fidoalliance.org/fido2/.