SOPG: Bincike-Bisa Tsarin Samar da Kalmar Sirri Mai Tsari don Cibiyoyin Jijiyoyi na Autoregressive

1. Gabatarwa

Kalmomin sirri sun ci gaba da zama babbar hanyar tabbatar da mai amfani saboda sauƙinsu da sassaucinsu. Saboda haka, zato kalmar sirri wani muhimmin sashi ne na binciken tsaron yanar gizo, wanda ke da mahimmanci ga duka gwajin tsaro na kai hari (misali, gwajin shiga, dawo da kalmar sirri) da kuma kimanta ƙarfin tsaro. Hanyoyin gargajiya, daga ƙididdiga na tushen ƙa'ida zuwa ƙirar ƙididdiga kamar sarƙoƙin Markov da PCFG, suna da iyakoki na asali a cikin bambance-bambance da inganci. Zuwan koyo mai zurfi, musamman cibiyoyin jijiyoyi na autoregressive kamar GPT, suna ba da hanya mai ban sha'awa don samar da ƙarin zato na kalmar sirri na gaskiya da inganci. Duk da haka, babban matsalar cikas ya ci gaba: daidaitaccen hanyar samfurin bazuwar yana haifar da sakamako masu maimaitawa kuma, mahimmanci, yana samar da kalmomin sirri a cikin tsari mara kyau, yana hana ingantaccen harin sosai. Wannan takarda ta gabatar da SOPG (Samuwar Kalmar Sirri Mai Tsari Bisa Bincike), wata sabuwar hanyar da aka ƙera don shawo kan wannan matsalar cikas.

2. Bayan Fage & Ayyukan Da Suka Danganta

3. Hanyar SOPG

4. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Ga model ɗin autoregressive, yuwuwar kalmar sirri $\mathbf{x} = (x_1, x_2, ..., x_T)$ an raba shi kamar haka: $$P(\mathbf{x}) = \prod_{t=1}^{T} P(x_t | x_{

5. Sakamakon Gwaji & Bincike

6. Tsarin Bincike & Misalin Lamari

Tsarin Kimanta Model ɗin Zato Kalmar Sirri:

1. Ƙarfin Ƙirƙira: Shin tsarin ginin zai iya koyon rarraba kalmomin sirri masu rikitarwa daidai? (misali, GPT da GAN).
2. Dabarar Samuwa: Yaya ake samun samfurin 'yan takara daga model? (Bazuwar da Tsari/Bisa bincike).
3. Ma'auni na Ingantaccen Hari:
   • Ƙimar Rufe: % na kalmomin sirri na gwaji da aka karya a cikin N zato.
   • Lambar Zato: Adadin zato da ake buƙata don karya X% na kalmomin sirri.
   • Ingantacciyar Ƙima: % na zato da aka samar waɗanda suke ingantattun kalmomin sirri na musamman.
   • Farashin Lissafi/Lokaci: Ƙididdiga ko lokaci kowane zato.

Misalin Lamari (Ba Code ba): Ka yi la'akari da masu kai hari guda biyu, Alice da Bob, suna amfani da irin wannan model ɗin PassGPT da aka horar. Alice tana amfani da daidaitaccen samfurin bazuwar. Bob yana amfani da hanyar SOPG da aka haɗa da PassGPT (wanda ya sa ya zama SOPGesGPT). Don karya 20% na jerin kalmar sirri da aka yi niyya, mai samfurin Alice na iya buƙatar samar da zato miliyan 5, tare da maimaitawa da yawa, yana ɗaukar sa'o'i 10. Tsarin Bob na tushen SOPG yana samar da kalmomin sirri a cikin tsari na yiwuwa, yana karya irin wannan 20% tare da zato na musamman 500,000 kawai, masu yuwuwa, yana kammala aikin a cikin sa'a 1. Harin Bob ya fi inganci sau 10 dangane da zato da lokaci, fa'ida mai mahimmanci.

7. Hangen Nesa na Aikace-aikace & Hanyoyin Gaba

Aikace-aikace Nan da Nan:

• Gwajin Ƙarfin Kalmar Sirri Mai Ƙarfafawa: Ƙungiyoyin tsaro za su iya amfani da model ɗin da aka haɓaka SOPG don bincika manufofin kalmar sirri da inganci, gano raunin kalmomin sirri kafin masu kai hari su yi.
• Binciken Dijital & Tilasta Bin Doka: Haɓaka dawo da kalmar sirri daga na'urorin da aka kwace a cikin binciken laifuka.
• Ingantattun Jerin Baƙaƙen Kalmar Sirri: Samar da cikakkun jerin kalmomin sirri masu rauni da tsari na yiwuwa don ƙin tsarin yayin ƙirƙira.

Hanyoyin Bincike na Gaba:

• Bincike Haɗaɗɗe & Mai Daidaitawa: Haɗa SOPG tare da wasu hasashe na bincike ko sanya shi ya dace bisa halayen manufa (misali, gidan yanar gizo, ƙididdiga na mai amfani).
• Tsaro akan Zato Mai Tsari: Bincike cikin sabbin tsare-tsaren hashing na kalmar sirri ko ka'idojin tabbatarwa waɗanda ke da juriya musamman ga hare-haren yiwuwa mai tsari, suna motsawa bayan tsaro na tushen entropy.
• Bayan Kalmomin Sirri: Yin amfani da ƙa'idodin samuwa mai tsari zuwa wasu yankuna na tsaro, kamar samar da maɓallan ɓoyewa masu yuwuwa ko tsarin kutsawa cibiyar sadarwa don gwaji.
• Haɓaka Inganci: Rage ƙwaƙwalwar ajiya da ƙarin lissafi na algorithm ɗin bincike don sanya shi ya zama mai girma ga mafi girman model da saitin haruffa.

8. Nassoshi

1. M. J. Weir et al., "Karya Kalmar Sirri Ta Amfani da Nahawun Mahallin Kyauta na Yiwuwa," a cikin IEEE Symposium on Tsaro da Keɓantawa, 2009.
2. B. Hitaj et al., "PassGAN: Hanyar Koyo Mai Zurfi don Zato Kalmar Sirri," a cikin Babban Taro na Duniya akan Tsaro na Cryptography da Tsaron Cibiyar Sadarwa, 2019.
3. J. Goodfellow et al., "Cibiyoyin Jijiyoyi Masu Gabaɗaya," a cikin Ci gaba a Cibiyoyin Sarrafa Bayanai na Jijiyoyi, 2014. (Takardar tushen GAN)
4. A. Vaswani et al., "Hankali Shine Abinda Kake Bukata," a cikin Ci gaba a Cibiyoyin Sarrafa Bayanai na Jijiyoyi, 2017. (Takardar tushen Transformer)
5. D. P. Kingma da M. Welling, "Karkatar da Kanta-Variational Bayes," arXiv:1312.6114, 2013. (Takardar tushen VAE)
6. M. Dell'Amico da P. Filippone, "Ƙimar Ƙarfi na Monte Carlo: Saurin Gwajin Kalmar Sirri Mai Dogaro," a cikin Babban Taro na ACM akan Tsaron Kwamfuta da Sadarwa, 2015.
7. OpenAI, "Rahoton Fasaha na GPT-4," 2023. (Yana nuna iyawar manyan model ɗin autoregressive).

9. Bincike na Asali & Sharhin Kwararru