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

1. Gabatarwa

Duk da ci gaban fasahar tabbatar da ainihi, kalmomin sirri sun kasance babbar hanyar saboda sauƙinsu da kuma yiwuwar amfani da su. Zubar da kalmomin sirri yana haifar da manyan barazanar tsaro, yana ba da damar shiga ba tare da izini ba da kuma haɓaka kayan aikin fasa kalmar sirri. Wannan takarda tana binciken aikace-aikacen Manyan Tsarin Harshe (LLMs) akan tsarin kalmar sirri, tare da gabatar da PassGPT—wani tsarin da aka horar da shi akan zubar da kalmomin sirri don samarwa da kimanta ƙarfi.

Binciken ya nuna cewa PassGPT ya fi hanyoyin da suka danganci Tsarin Gargadi na Halitta (GAN) girma ta hanyar zato kalmomin sirri 20% fiye da waɗanda ba a taɓa ganinsu ba kuma ya gabatar da samar da kalmar sirri mai jagora—wata sabuwar ƙwarewa don samar da kalmomin sirri a ƙarƙashin ƙayyadaddun sharuɗɗa.

2. Hanyoyi & Tsarin Gine-gine

An gina PassGPT akan tsarin gine-ginen GPT-2, wanda aka daidaita shi don samar da haruffan kalmar sirri a jere. Wannan hanyar ta bambanta da GANs waɗanda ke samar da kalmomin sirri a matsayin cikakkun raka'a.

2.1. Tsarin Ƙirar PassGPT

Tsarin wani Transformer ne mai cin gashin kansa wanda aka horar da shi akan manyan zubar da kalmomin sirri. Yana koyon rarraba yiwuwa $P(x_t | x_{

2.2. Samar da Kalmar Sirri Mai Jagora

Wani sabon abu mai mahimmanci shi ne samarwa mai jagora a matakin harafi. Ta hanyar sarrafa tsarin samfurin (misali, ta amfani da yiwuwar sharadi ko rufe fuska), PassGPT na iya samar da kalmomin sirri waɗanda suka gamsar da takamaiman ƙayyadaddun sharuɗɗa, kamar ƙunsar wasu alamomi, cika buƙatun tsayi, ko haɗa da takamaiman ɓangarori—wani aikin da ba za a iya cimma shi tare da GANs na yau da kullun ba.

2.3. Haɓaka PassVQT
PassVQT ya haɗa da dabarun Vector Quantized Transformer (VQT), ta amfani da littafin lambobi na daban don wakiltar abubuwan da aka ɓoye. Wannan na iya ƙara rudani da bambancin kalmomin sirri da aka samar, ko da yake yana iya zuwa da farashin lissafi.

3. Sakamakon Gwaji

3.1. Aikin Zato Kalmar Sirri

Gwaje-gwaje akan zubar da kalmomin sirri na ainihi (misali, RockYou) sun nuna PassGPT ya fi na baya-bayan nan mafi kyawun tsarin samarwa mai zurfi kamar PassGan girma. A cikin gwaji ɗaya, PassGPT ya zato kalmomin sirri na musamman, waɗanda ba a taɓa ganinsu ba sau biyu idan aka kwatanta da hanyoyin da suka danganci GAN. Haka kuma ya nuna ƙarfin haɓakawa zuwa sabbin bayanan da aka ajiye.

Kwatanta Aiki

PassGPT vs. GANs: Mafi girman nasarar kashi 20% wajen zato kalmomin sirri da ba a gani ba.

Haɓakawa: Ingantaccen aiki akan sabbin zubar da kalmomin sirri da ba a gani ba yayin horo.

3.2. Binciken Rarraba Yiwuwa

Ba kamar GANs ba, PassGPT yana ba da bayyanannen rarraba yiwuwa akan kalmomin sirri. Bincike ya nuna alaƙa mai ƙarfi tsakanin ƙarancin yuwuwar kalmar sirri (babban mummunan log-yiwuwa) da babban ƙarfi kamar yadda masu kimanta kamar zxcvbn suka auna. Duk da haka, PassGPT ya gano lokuta inda kalmomin sirri da masu kimanta na al'ada suka ɗauka "mai ƙarfi" suna da yuwuwa mai yawa a ƙarƙashin tsarinsa, yana nuna yuwuwar raunin tsaro.

Ma'anar Jadawali: Zane mai zato zai nuna yuwuwar kalmar sirri (PassGPT) akan ginshiƙin x da maki ƙarfi (zxcvbn) akan ginshiƙin y, yana bayyana wani yanayi na gaba ɗaya mara kyau tare da fitattun abubuwan da ba a saba gani ba inda kalmomin sirri masu ƙarfi suna da yuwuwa mai yawa da ba a zata ba.

4. Binciken Fasaha & Tsarin Aiki

Hangen Nesa na Mai Binciken Masana'antu: Ƙima mai mahimmanci game da hanyar PassGPT, abubuwan da ke tattare da ita, da abubuwan da za a iya amfani da su a aikace.

4.1. Cikakken Fahimta

Babban nasarar takardar ba wani ƙarin tsarin AI ne kawai don kalmomin sirri ba; canji ne daga daban-daban tsarin daidaitawa zuwa samar da tsarin jeri. Yayin da kayan aiki kamar Hashcat suka dogara da ƙa'idodi da sarƙoƙin Markov, kuma GANs kamar PassGAN suna samar da cikakkun sakamako, PassGPT yana ɗaukar ƙirƙirar kalmar sirri a matsayin aikin harshe. Wannan yana kwatanta yadda LLMs kamar GPT-3 suka kama "nahawu" da "ma'anar" harshe na halitta, amma an yi amfani da su a nan ga "harshen" ƙirƙirar kalmar sirri na ɗan adam. Ainihin ƙimar da ke bayyana ita ce bayyanannen, rarraba yiwuwa mai iya sarrafawa da yake bayarwa—wani siffa da ba a cika ganinta a cikin GANs ba, waɗanda galibi ake suka da cewa "akwatunan baƙi" (Goodfellow et al., 2014). Wannan yana motsa tsaron kalmar sirri daga zato na hasashe zuwa tunani na yiwuwa.

4.2. Tsarin Ma'ana

Hujjar ta ci gaba da ma'ana mai ƙarfi: (1) LLMs sun mamaye NLP ta hanyar ƙirar jerin gwano; (2) kalmomin sirri jerin haruffa ne tare da tsarin ɓoye; (3) saboda haka, LLMs yakamata su yi tsarin kalmomin sirri yadda ya kamata. Tabbatarwa tana da ƙarfi: mafi girman aikin zato ya tabbatar da jigon. Gabatar da samarwa mai jagora wani faɗaɗa ne na halitta na tsarin jeri—kamar samar da rubutu mai sarrafawa a cikin tsarin kamar CTRL (Keskar et al., 2019). Binciken rarraba yiwuwa shine muhimmin mataki na gaba, yana haɗa samar da ƙira zuwa ainihin yanki na kimanta ƙarfi. Gudu daga ƙira -> samarwa -> bincike -> aikace-aikace yana da haɗin kai kuma yana da tasiri.

4.3. Ƙarfi & Kurakurai

Ƙarfi: Ribar aikin ba za a iya musantawa ba. Ƙarfin samarwa mai jagora sabon abu ne na gaske tare da aikace-aikace nan take don gwajin shiga ciki (samar da ƴan takarar kalmar sirri masu bin ƙa'ida) kuma mai yiwuwa don taimaka wa masu amfani su ƙirƙiri kalmomin sirri masu tunawa amma masu rikitarwa. Bayar da rarraba yiwuwa babbar fa'ida ce ta ka'ida da aikace, yana ba da damar lissafin entropy da haɗawa tare da tsarin tsaro na yanzu.

Kurakurai & Damuwa: Takardar ta yi watsi da manyan batutuwa. Na farko, amfani mai ma'ana biyu: Wannan kayan aiki ne mai ƙarfi na fasa kalmar sirri. Yayin da aka sanya shi don binciken "zato na layi", yuwuwar amfani da shi ba daidai ba tana da yawa, kuma sakin lambar/tsarin yana buƙatar ƙa'idodin ɗabi'a masu tsauri, kamar muhawarar da ke tattare da sauran binciken AI mai amfani biyu (Brundage et al., 2018). Na biyu, dogaro da bayanai: Kamar duk tsarin ML, PassGPT yana da kyau kamar yadda bayanan horonsa suke. Yana iya kasa yin tsarin kalmomin sirri daga al'adu ko harsuna waɗanda ba a wakilta su yadda ya kamata a cikin zubar da kowa. Na uku, farashin lissafi: Horarwa da gudanar da manyan masu canzawa yana da tsada idan aka kwatanta da wasu tsofaffin hanyoyi, wanda zai iya iyakance aikace-aikace na ainihin lokaci. An ambaci ƙarin "rudani" na nau'in PassVQT amma ba a tantance shi sosai ba—shin ƙarin bambance-bambance yana fassara zuwa zato mai inganci, ko kuma kawai ƙarin kirtani marasa ma'ana?

4.4. Abubuwan Da Za'a Iya Aiwatarwa

Ga Ƙungiyoyin Tsaro: Nan da nan ku tantance yadda ƙa'idodin kalmar sirri na ƙungiyar ku na iya zama maras ƙarfi ga wannan sabon tsarin harin da AI ke jagoranta. Manufofin da ke tilasta rikitattun tsari amma masu hasashe (misali, "Sunan Kamfani2024!") yanzu sun fi fallasa. Yi kira don canzawa zuwa amfani da bazuwar gaskiya (manajojin kalmar sirri) ko jimlolin sirri.

Ga Masu Bincike & Masu Sayarwa: Haɗa ƙididdiga na yiwuwar da suka danganci LLM cikin ma'aunin ƙarfi. Na'urar kimanta gauraye wanda ya haɗa ƙa'idodin gargajiya (zxcvbn) tare da yuwuwar PassGPT na iya zama mai ƙarfi. Haɓaka tsarin tsaro waɗanda za su iya gano kalmomin sirri da PassGPT zai iya samarwa, ƙirƙirar tsarin AI da ya danganci tsarin yaƙi a cikin tsaron kalmar sirri.

Ga Masu Tsara Manufofi: Ku ba da kuɗin bincike zuwa aikace-aikacen tsaro na wannan fasaha kuma ku kafa tsarin ɗabi'a bayyananne don buga ƙwararrun kayan aikin AI na kai hari a cikin tsaron yanar gizo.

Misalin Tsarin Aiki (Ba Lamba ba): Yi la'akari da manufar kalmar sirri na cibiyar kuɗi: "Haruffa 12, 1 na sama, 1 na ƙasa, 1 lamba, 1 harafi na musamman." Tsohuwar kayan aikin fasa kalmar sirri na iya amfani da ƙarfi ko amfani da ƙa'idodin ɓarna. GAN na iya fuskantar wahalar samar da sakamako waɗanda suka cika duk waɗannan ƙayyadaddun sharuɗɗan. Samarwa mai jagora na PassGPT za a iya jagorantar shi don ɗaukar samfuri kawai jerin gwano waɗanda suka cika wannan takamaiman manufa, yana bincika yankin mai yuwuwa mai yawa na wannan sararin bincike mai ƙayyadaddun sharuɗɗa, yana mai da shi kayan aiki mai ƙarfi ga duka ƙungiyoyin ja masu gwada wannan manufa da kuma masu kai hari na akwatin baƙi.

5. Aikace-aikace na Gaba & Jagorori

Haɓaka Ƙimar Ƙarfi: Haɗa makin yiwuwar PassGPT cikin ma'aunin ƙarfin kalmar sirri na ainihin lokaci don gidajen yanar gizo da aikace-aikace.
Binciken Kalmar Sirri Mai Tsari: Ƙungiyoyi na iya amfani da tsarin PassGPT mai jagora don ƙirƙira da gwada kalmomin sirri waɗanda suka bi manufofin cikin gida, gano raunin kafin masu kai hari su yi.
Tsarin Tsaro na Gauraye: Haɓaka tsarin daban-daban waɗanda za su iya bambanta tsakanin kalmomin sirri da mutum ya zaɓa da waɗanda LLM ya samar don alamar takaddun shaida masu yuwuwar lalacewa ko rauni.
Ƙirar Jerin Gwano na Yankuna Daban-daban: Yin amfani da irin wannan tsarin gine-gine zuwa wasu jerin gwano masu mahimmanci ga tsaro, kamar alamun yatsa na yarjejeniyar sadarwa, jerin kiran API na malware, ko tsarin ma'amala na zamba.
Horo na Tarayya & Kiyaye Sirri: Bincika dabarun horar da irin waɗannan tsarin akan rarraba, bayanan kalmar sirri marasa suna ba tare da tattara zubar da bayanai masu mahimmanci ba.
Samar da Kalmar Sirri na Adawa: Yin amfani da samarwa mai jagora don ƙirƙirar "misalai na adawa"—kalmomin sirri waɗanda suke bayyana masu ƙarfi ga masu kimanta amma cikin sauƙi PassGPT yana iya zato su—don gwadawa da ƙarfi da haɓaka waɗannan masu kimanta.

6. Nassoshi

Rando, J., Perez-Cruz, F., & Hitaj, B. (2023). PassGPT: Password Modeling and (Guided) Generation with Large Language Models. arXiv preprint arXiv:2306.01545.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., ... & Bengio, Y. (2014). Generative adversarial nets. Advances in neural information processing systems, 27.
Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language models are unsupervised multitask learners. OpenAI blog, 1(8), 9.
Hitaj, B., Gasti, P., Ateniese, G., & Perez-Cruz, F. (2019). PassGAN: A Deep Learning Approach for Password Guessing. In Applied Cryptography and Network Security.
Keskar, N. S., McCann, B., Varshney, L. R., Xiong, C., & Socher, R. (2019). Ctrl: A conditional transformer language model for controllable generation. arXiv preprint arXiv:1909.05858.
Brundage, M., Avin, S., Clark, J., Toner, H., Eckersley, P., Garfinkel, B., ... & Amodei, D. (2018). The malicious use of artificial intelligence: Forecasting, prevention, and mitigation. arXiv preprint arXiv:1802.07228.
Wheeler, D. L. (2016). zxcvbn: Low-budget password strength estimation. In USENIX Security Symposium.