Koyon Zurfin AI don Ƙirƙirar Kalmar Sirri: Nazarin Kwatance

1. Gabatarwa da Dalili

Tabatar da asali ta hanyar kalmar sirri na ci gaba da zama ruwan dare saboda sauƙinta da sanin mai amfani. Duk da haka, kalmomin sirri da masu amfani suka zaɓa sau da yawa ana iya hasashensu, gajere, kuma ana sake amfani da su a duk faɗin dandamali, suna haifar da manyan raunin tsaro. Wannan takarda tana bincika ko tsarin koyon zurfin AI na iya koyo da kuma kwaikwayon waɗannan tsarin ƙirƙirar kalmar sirri na ɗan adam don samar da ƙwaƙƙwaran ƴan takarar kalmar sirri don gwajin tsaro da bincike.

Canji daga hasashen kalmar sirri na tushen ƙa'ida, wanda ƙwararru ke jagoranta (misali, tsarin Markov, nahawu na mahallin kyauta) zuwa hanyoyin koyon zurfin AI na tushen bayanai kawai yana wakiltar canjin tsari. Wannan aikin yana bincika tarin tsare-tsare mai faɗi, gami da hanyoyin hankali, masu yin kankare, da cibiyoyin adawa na ƙirƙira, tare da sabon gudunmawa wajen amfani da Masu Yin Kankare na Bambance-bambance (VAEs) a wannan fanni.

2. Ayyukan Da Aka Yi da Baya

Hasashen kalmar sirri na gargajiya ya dogara ne akan nazarin ƙididdiga na bayanan da aka yi fashi (misali, RockYou) don ƙirƙirar tsarin ƙa'ida da tsarin ƙima kamar sarƙoƙin Markov. Waɗannan hanyoyin suna buƙatar ƙwarewar fanni don ƙirƙirar ƙa'idodi masu inganci. Sabanin haka, koyon zurfin AI na zamani don ƙirƙirar rubutu, wanda tsarin gine-gine kamar Transformer (Vaswani da sauransu, 2017) da ci gaban horo suka haɗa, yana koyon tsari kai tsaye daga bayanai ba tare da ƙirƙirar ƙa'ida a sarari ba.

Manyan ci gaban da ke ba da damar wannan binciken sun haɗa da:

• Hanyoyin Hankali: Tsare-tsare kamar BERT da GPT suna ɗaukar rikice-rikicen alaƙar mahalli a cikin jerin bayanai.
• Koyon Wakilci: Masu yin kankare suna koyon matsanancin wakilci mai ma'ana (sararin ɓoye) na bayanai.
• Horo Mai Zurfi: Dabarun kamar ƙididdiga na bambance-bambance da daidaitawar Wasserstein suna daidaitawa da inganta horon tsarin ƙirƙira.

3. Tsarin Koyon Zurfin AI na Ƙirƙira

Wannan sashe yana cikakken bayani game da ainihin tsare-tsaren da aka kimanta don ƙirƙirar kalmar sirri.

3.1 Cibiyoyin Hankali na Jijiya

Tsare-tsaren da ke amfani da tsarin kula da kai ko Transformer suna sarrafa kirtan kalmar sirri a matsayin jerin haruffa ko alamun. Tsarin hankali yana ba da damar tsarin auna muhimmancin haruffa daban-daban a cikin mahallin, yana koyon ginshiƙan ginshiƙai na gama-gari (kamar "123" ko "password") da kuma sanya su.

3.2 Hanyoyin Yin Kankare

Masu yin kankare na yau da kullun suna matsawa kalmar sirri ta shigarwa zuwa cikin vector ɓoye kuma suna ƙoƙarin sake gina ta. Matsalolin toshewa suna tilasta wa tsarin koyon mahimman siffofi. Duk da yake yana da amfani don wakilci, masu yin kankare na yau da kullun ba su da asalin ƙirƙira don sabbin samfura.

3.3 Cibiyoyin Adawa na Ƙirƙira (GANs)

GANs suna sanya cibiyar ƙirƙira (ƙirƙirar kalmomin sirri) a kan cibiyar nema (yanke hukunci na gaskiya). Ta hanyar horon adawa, mai ƙirƙira yana koyon samar da samfuran da ba za a iya bambanta su da kalmomin sirri na gaske ba. Duk da haka, GANs suna da wahalar horarwa kuma suna iya fuskantar rugujewar yanayi, inda suke samar da iyakacin iri-iri.

3.4 Masu Yin Kankare na Bambance-bambance (VAEs)

Babban gudunmawar wannan aikin shine amfani da VAEs. Ba kamar masu yin kankare na yau da kullun ba, VAEs suna koyon sararin ɓoye na ƙima. Mai ɓoyewa yana fitar da sigogi (ma'ana $\mu$ da bambanci $\sigma^2$) na rarraba Gaussian. Ana samun samfurin vector ɓoye $z$: $z \sim \mathcal{N}(\mu, \sigma^2)$. Sannan mai ɓoyewa yana sake gina shigarwar daga $z$.

Ayyukan asara shine Ƙananan Shaida (ELBO):

$\mathcal{L}_{VAE} = \mathbb{E}_{q_{\phi}(z|x)}[\log p_{\theta}(x|z)] - D_{KL}(q_{\phi}(z|x) \| p(z))$

Kalma ta farko ita ce asarar sake gini. Kalma ta biyu, bambancin Kullback-Leibler, tana daidaita sararin ɓoye don kusanci da rarraba da ta gabata $p(z)$ (yawanci al'ada ta al'ada). Wannan tsarin sararin ɓoye yana ba da damar siffofi biyu masu ƙarfi don hasashen kalmar sirri:

1. Haɗaɗɗiya: Samfurin maki tsakanin vectors ɓoye biyu na sanannun kalmomin sirri na iya haifar da sabbin kalmomin sirri masu haɗaka waɗanda ke haɗa siffofi na duka biyun.
2. Samfurin Da Aka Yi niyya: Ta hanyar daidaita sararin ɓoye ko bincika a cikinsa, mutum na iya samar da kalmomin sirri tare da takamaiman kaddarorin (misali, ya ƙunshi wani ɓangaren rubutu).

4. Tsarin Gwaji da Bayanan Gwaji

Binciken yana amfani da tsari ɗaya, mai sarrafawa don kwatanta gaskiya. An horar da tsare-tsare kuma an kimanta su akan bayanan fashi na kalmar sirri na duniya da yawa da aka sani:

• RockYou: Babban bayanai na gargajiya daga keta aikace-aikacen zamantakewa.
• LinkedIn: Kalmomin sirri daga keta hanyar sadarwar ƙwararru, galibi ana tunanin sun fi rikitarwa.
• Youku, Zomato, Pwnd: Ƙarin bayanan daga ayyuka daban-daban waɗanda ke ba da bambancin salo na kalmar sirri da tasirin al'adu.

Ma'aunin kimantawa sun haɗa da:

• Adadin Daidaito: Kashi na kalmomin sirri da aka ƙirƙira waɗanda suka yi nasara suka dace da kalmomin sirri a cikin saitin gwaji da aka keɓe (kwaikwayon ƙoƙarin fashewa).
• Keɓancewa: Kashi na kalmomin sirri da aka ƙirƙira waɗanda suka bambanta da juna.
• Sabon Abu: Kashi na kalmomin sirri da aka ƙirƙira waɗanda ba a samu su a cikin bayanan horo ba.

5. Sakamako da Nazarin Aiki

Nazarin ƙwaƙƙwaran ya bayyana yanayin aiki mai zurfi:

• VAEs Sun Fito a matsayin Mai Aiki Mai Ƙarfi: Tsarin VAE da aka tsara sun cimma mafi kyawun ko gasa sosai adadin daidaito a duk faɗin bayanan. Tsarin sararin ɓoye nasu yana ba da fa'ida mai mahimmanci wajen samar da samfura daban-daban da masu ma'ana, wanda ke haifar da babban maki na keɓancewa da sabon abu.
• GANs Suna Nuna Babban Ƙarfinsu Amma Rashin Kwanciyar Hankali: Lokacin da aka horar da su cikin nasara, GANs na iya samar da kalmomin sirri masu gaskiya sosai. Duk da haka, aikin su bai da daidaituwa, sau da yawa suna fama da rugujewar yanayi (ƙananan keɓancewa) ko kuma sun kasa haɗuwa, suna daidaitawa da ƙalubalen horon GAN da aka rubuta a cikin takarda ta asali ta Goodfellow da sauransu da kuma bincike na baya kamar "Wasserstein GAN" na Arjovsky da sauransu.
• Tsarin Hankali Suna Ƙware wajen Kama Tsarin Gida: Tsare-tsare kamar tsarin Transformer suna da tasiri sosai wajen koyon n-grams na haruffa na gama-gari da dogaro da matsayi (misali, babban harafin farko, ƙara lambobi a ƙarshe).
• Bambancin Bayanan Yana Da Muhimmanci: Matsayin aikin tsari na iya canzawa dangane da bayanan. Misali, tsare-tsaren da suke aiki da kyau akan RockYou ƙila ba za su yi aiki da kyau ga LinkedIn ba, wanda ke nuna mahimmancin bambancin bayanan horo.

Fassarar Chati (Hasashen bisa bayanin takarda): Chati na sandar da ke kwatanta tsare-tsare zai iya nuna VAEs da manyan tsare-tsaren Hankali suna jagoranci a cikin adadin daidaito. Hoton zane na Keɓancewa da Adadin Daidaito zai nuna VAEs a cikin madaidaicin yanki (babba akan duka axes), yayin da wasu lokutan GAN na iya taruwa a cikin yanki mai babban adadin daidaito amma ƙananan keɓancewa, yana nuna rugujewar yanayi.

6. Nazarin Fasaha da Fahimta

Babban Fahimta

Mafi ƙarfin fahimtar takardar ita ce ƙirƙirar kalmar sirri ba kawai matsalar ƙirar jerin bayanai ba ce; matsala ce ta ƙididdige yawa a cikin sararin ɓoye mai tsari. Yayin da RNNs/Transformers suka ƙware wajen hasashen harafi na gaba, ba su da bayyanannen, tsarin kewayawa na "manifold na kalmar sirri." VAEs suna ba da wannan ta hanyar ƙira. Marubutan sun gano daidai cewa ikon yin samfurin da aka yi niyya (misali, "ƙirƙiri kalmomin sirri masu kama da wannan ƙa'idar sunan kamfani") da haɗaɗɗiyar sassauƙa tsakanin nau'ikan kalmar sirri shine mai canza wasa don binciken tsaro na tsari, wucewa fiye da ƙididdigar ƙarfi.

Kwararar Hankali

Hankalin binciken yana da inganci: 1) Tsara hasashen kalmar sirri a matsayin aikin ƙirƙirar rubutu. 2) Aiwatar da kayan aikin DL na zamani (Hankali, GANs, VAEs). 3) Mafi mahimmanci, gane cewa kaddarorin sararin ɓoye na VAEs suna ba da fa'idodin aiki na musamman akan sauran tsare-tsaren ƙirƙira. 4) Tabbatar da wannan hasashe ta hanyar ingantaccen ƙididdiga, mai yawan bayanai. Kwararar daga daidaita tsari zuwa tabbacin ƙwaƙƙwaran bayanai yana bayyana kuma yana jan hankali.

Ƙarfi & Kurakurai

Ƙarfi: Tsarin kwatancen babban ƙarfi ne. Sau da yawa, takardu suna gabatar da tsari guda ɗaya. Anan, ƙididdiga akan GANs da tsarin hankali yana ba da mahallin mahimmanci, yana nuna VAEs ba kawai sun bambanta ba, amma suna ba da mafi kyawun ciniki tsakanin ingancin samfur, bambancin, da kuma sarrafawa. Mayar da hankali kan bayanan duniya na gaske (LinkedIn, Zomato) ya kafa binciken a cikin gaskiyar aiki.

Kurakurai: Takardar, kamar yawancin fannin, tana aiki a cikin tsarin bayan keta. Tana nazarin alamun (kalmomin sirri da aka yi fashi) maimakon cutar (kanta tabbatar da asali ta kalmar sirri). An yarda da "takobi mai kaifi biyu" na ɗa'a amma ba a bincika shi sosai ba. Bugu da ƙari, yayin da VAEs suka inganta sarrafawa, tsarin samfurin har yanzu yana da ƙasa kai tsaye fiye da tsarin tushen ƙa'ida don mai binciken ɗan adam. "Ma'anoni" na sararin ɓoye, duk da yana da tsari, na iya zama marar haske.

Fahimta Mai Aiki

Ga ƙungiyoyin tsaro: Haɗa masu ƙirƙira na tushen VAE cikin kayan aikin ku na binciken kalmar sirri na gaggawa. Siffar samfurin da aka yi niyya ita ce mabuɗin ƙirƙirar lissafin kalmomi na musamman don gwaje-gwajen shiga kan takamaiman ƙungiyoyi ko ƙididdiga na masu amfani.

Ga masu ƙirƙirar manufofin kalmar sirri: Waɗannan tsare-tsaren suna zama kankara mai nuna iyakokin halayen ɗan adam da ake iya hasashensu. Idan VAE zai iya hasashensa, ba kalmar sirri mai kyau ba ce. Dole ne manufofin su tilasta gaskiyar bazuwar ko amfani da jumlar shiga, wucewa fiye da ƙa'idodin tsarawa waɗanda waɗannan tsare-tsaren ke koyi cikin sauƙi.

Ga masu binciken AI: Wannan aikin shiri ne don amfani da tsare-tsaren ƙirƙira masu tsari (VAEs, Kwararar Daidaitawa) zuwa wasu matsalolin tsaro na jerin bayanai masu rarrabuwa, kamar ƙirƙirar sa hannun malware ko kwaikwayon zirga-zirgar cibiyar sadarwa. Dabarun binciken sararin ɓoye ana iya canza su kai tsaye.

Misalin Tsarin Nazari

Yanayi: Kamfanin tsaro yana bincika kamfani inda ake zargin kalmomin sirri na ma'aikata sun dogara ne akan sunan aikin "ProjectPhoenix" da shekara "2023".

Hanyar Tushen Ƙa'ida ta Gargajiya: Ƙirƙiri ƙa'idodin hannu: {ProjectPhoenix, phoenix, PHOENIX} + {2023, 23, @2023} + {!, #, $}. Wannan yana ɗaukar lokaci mai yawa kuma yana iya rasa bambance-bambancen ƙirƙira.

Hanyar Ingantaccen VAE:

1. Ƙididdige sanannun kalmomin sirri masu rauni (misali, "ProjectPhoenix2023", "phoenix23") zuwa cikin sararin ɓoye na VAE.
2. Yi tafiya kai tsaye ko samfurin a cikin yankin ɓoye a kusa da waɗannan maki, wanda aka jagoranta ta hanyar rarraba abubuwan gama gari da aka koya, musanyawar leetspeak, da tsarin babban haruffa.
3. Ƙididdige vectors ɓoye da aka samu don samar da lissafin kalmomi da aka yi niyya: misali, "pr0jectPh0enix#23", "PH0ENIX2023!", "project_phoenix23".

Wannan hanyar tana bincika sararin bambance-bambancen da za a iya yiwuwa da bayanan horo suka nuna, mai yiwuwa ya gano kalmomin sirri da marubucin ƙa'idar ɗan adam ba zai yi tunani ba.

7. Ayyuka na Gaba da Jagorori

Hanyar wannan binciken tana nuni zuwa ga manyan jagorori na gaba da yawa:

1. Tsare-tsaren Haɗakaɗɗu & Daidaitattun: Tsare-tsaren gaba mai yiwuwa za su haɗa ƙarfin gine-gine daban-daban—misali, amfani da Transformer a matsayin mai ɓoyewa/mai ɓoyewa a cikin tsarin VAE, ko daidaita GANs/VAEs akan bayanan taimako kamar ƙididdiga na masu amfani (wanda aka ƙaddara daga wasu ketare) ko nau'in gidan yanar gizo don samar da ƴan takara mafi yawan niyya.
2. Kariya ta Gaggawa & Mitocin Ƙarfin Kalmar Sirri: Aikace-aikacen mafi ɗa'a da tasiri shine jujjuya rubutun. Waɗannan tsare-tsaren ƙirƙira na iya ƙarfafa ƙarni na gaba na masu ƙididdige ƙarfin kalmar sirri. Maimakon duba ƙamus masu sauƙi, mita na iya amfani da tsarin ƙirƙira don ƙoƙarin hasashen kalmar sirri a lokacin gaskiya kuma ya ba da maki mai ƙarfi na ƙarfi dangane da yadda aka samar da shi cikin sauƙi.
3. Bayan Kalmomin Sirri: Hanyoyin suna amfani da su kai tsaye ga wasu fannonin tsaro waɗanda ke buƙatar samar da bayanan gaskiya, masu tsari masu rarrabuwa: samar da imel ɗin satar bayanai na roba, ƙirƙirar zirga-zirgar cibiyar sadarwa na yaudara, ko kwaikwayon halayen mai amfani don tsarin kudan zuma.
4. Ƙarfin Adawa: Yayin da waɗannan masu ƙirƙira suka inganta, za su tilasta haɓaka ingantaccen tabbatar da asali. Bincike don ƙirƙirar kalmomin sirri waɗanda suke ƙarfi ga adawa da waɗannan masu hasashen AI—kalmomin sirri waɗanda ke tunawa ga mutane amma suna kwance a cikin yankunan sararin ɓoye waɗanda tsarin ya ba da ƙima kaɗan—zai iya zama sabon yanki.

8. Nassoshi

1. Biesner, D., Cvejoski, K., Georgiev, B., Sifa, R., & Krupicka, E. (2020). Generative Deep Learning Techniques for Password Generation. arXiv preprint arXiv:2012.05685.
2. 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.
3. Kingma, D. P., & Welling, M. (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114.
4. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30.
5. Arjovsky, M., Chintala, S., & Bottou, L. (2017). Wasserstein generative adversarial networks. International conference on machine learning (pp. 214-223). PMLR.
6. Weir, M., Aggarwal, S., Medeiros, B., & Glodek, B. (2009). Password cracking using probabilistic context-free grammars. 2009 30th IEEE Symposium on Security and Privacy (pp. 391-405). IEEE.
7. National Institute of Standards and Technology (NIST). (2017). Digital Identity Guidelines (SP 800-63B).