SOPG: Bincike-Bisa Tsarin Samar da Kalmar Sirri Mai Tsari don Cibiyoyin Sadarwar Jijiyoyi Masu Gudanar da Kai (Autoregressive Neural Networks)

Teburin Abubuwan Ciki

1. Gabatarwa

Kalmomin sirri sun kasance mafi yawan hanyar tabbatar da asalin mai amfani. Saboda haka, zato na kalmar sirri wani muhimmin bangare ne na binciken tsaron yanar gizo, wanda ke goyan bayan duka gwajin tsaro na kai hari (fasa) da kuma kimanta ƙarfin tsaro. Hanyoyin gargajiya, daga ƙididdiga na tushen ƙa'ida zuwa tsarin ƙididdiga kamar sarƙoƙin Markov da PCFG, suna da iyakoki na asali a cikin inganci da bambancin. Zuwan koyo mai zurfi, musamman cibiyoyin sadarwar jijiyoyi masu gudanar da kai (autoregressive), sun yi alƙawarin canza tsari. Duk da haka, wata matsala mai mahimmanci ta ci gaba: daidaitaccen hanyar samar da samfurin bazuwa. Wannan yana haifar da maimaita kalmomin sirri kuma, mafi muni, tsari na bazuwa na samarwa, yana tilasta wa maharan su tace manyan jerin abubuwa marasa inganci. Wannan takarda ta gabatar da SOPG (Samar da Kalmar Sirri Mai Tsari Bisa Bincike), wata sabuwar hanya da aka ƙera don sa ƙirar zato na kalmar sirri masu gudanar da kai su samar da kalmomin sirri a cikin kusan tsari mai saukowa na yuwuwar, ta haka yana ƙara ingantaccen kai hari sosai.

2. Bayan Fage & Ayyukan Da Suka Danganta

2.1 Juyin Halittar Zato na Kalmar Sirri

Zato na kalmar sirri ya samo asali ta hanyar matakai daban-daban. Hanyoyin farko sun dogara da harin ƙamus da ƙirar ƙa'idodin gurɓatawa da hannu (misali, John the Ripper), waɗanda suka kasance na gwargwado da dogaro da gogewa. Yaduwar ɓarkewar kalmomin sirri masu girma (misali, RockYou a 2009) ya ba da damar hanyoyin ƙididdiga masu dogaro da bayanai. Tsarin Markov (Weir et al., 2009) da Ƙa'idar Nahawu Mai Yuwuwar Ba tare da Mahallin Ba (PCFG) (Ma et al., 2014) sun ba da tsarin mafi tsari, na tushen yuwuwar don samarwa, ko da yake suna da haɗarin wuce gona da iri kuma ba su da ikon ƙirar rikitattun dogaro mai nisa a cikin tsarin kalmar sirri.

2.2 Hanyoyin Cibiyar Sadarwar Jijiyoyi

Ƙirar koyo mai zurfi, musamman Cibiyoyin Sadarwar Jijiyoyi Masu Gaba da Juna (GANs) kamar PassGAN (Hitaj et al., 2017) da ƙirar masu gudanar da kai (autoregressive) kamar waɗanda suka dogara da tsarin LSTM ko GPT, suna koyon rarraba yuwuwar kalmomin sirri kai tsaye daga bayanai. Suna iya samar da kalmomin sirri masu bambancin gaske da na gaskiya. Duk da haka, yawanci suna amfani da zato na bazuwa (misali, zato mai yawa) daga rarraba da aka koya a kowane mataki na samarwa. Wannan tsari na asali bai san matsayi na duniya na cikakkun yuwuwowin kalmar sirri ba, yana haifar da rashin ingancin da SOPG ke nufin warwarewa.

3. Hanyar SOPG

3.1 Babban Manufa

SOPG yana sake tsara samar da kalmar sirri daga matsalar zato zuwa matsalar bincike mai jagora. Maimakon zaɓar harafi na gaba da bazuwa, yana amfani da algorithm na bincike (mai yuwuwar bambancin binciken katako ko binciken mafi kyau) don bincika sararin ci gaba mai yuwuwar kalmar sirri, yana ba da fifiko ga hanyoyin da ke kaiwa ga cikakkun kalmomin sirri tare da mafi girman yuwuwar da aka ƙiyasta. Manufar ita ce fitar da jerin kalmar sirri a cikin tsari wanda ya kusanci daidaitaccen tsari mai saukowa ta $P(kalmar sirri|ƙira)$.

3.2 Algorithm na Bincike

Duk da yake taƙaitaccen PDF bai yi cikakken bayani game da takamaiman algorithm ba, halin da aka bayyana yana nuna hanyar da ke kiyaye jerin ginshiƙai na gabaɗaya na gurbin kalmar sirri. A kowane mataki, yana faɗaɗa gurbi mafi ban sha'awa (mafi girman yuwuwar tarawa) ta hanyar tambayar cibiyar sadarwar jijiyoyi don rarraba harafi na gaba, samar da sababbin ƴan takara. Ta hanyar bincika yankuna masu yuwuwar sararin kalmar sirri da farko, yana tabbatar da samar da farko na mafi yuwuwar kalmomin sirri kuma a asalinsa yana guje wa maimaitawa.

3.3 Tsarin SOPGesGPT

Marubutan sun aiwatar da hanyarsu akan tsarin tushen GPT, suna ƙirƙirar SOPGesGPT. Tsarin GPT (misali, mai fassara kawai) ana horar da shi akan bayanan kalmomin sirri da suka ɓace don hasashen harafi na gaba a cikin jerin. Ana amfani da SOPG a matsayin hanyar samarwa/ƙididdiga a saman wannan ƙirar da aka horar, yana maye gurbin daidaitaccen samfurin.

4. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Ƙirar mai gudanar da kai tana ayyana yuwuwar kalmar sirri $\mathbf{x} = (x_1, x_2, ..., x_T)$ a matsayin samfurin yuwuwowin sharadi: $$P(\mathbf{x}) = \prod_{t=1}^{T} P(x_t | x_1, ..., x_{t-1})$$ inda $x_t$ shine harafin a matsayi $t$, kuma $T$ shine tsawon kalmar sirri. Daidaitaccen samfurin yana zaɓar $x_t \sim P(\cdot | x_1, ..., x_{t-1})$.

SOPG, a zahiri, yana nufin nemo da fitar da jerin abubuwa $\mathbf{x}$ cikin tsari na raguwa $P(\mathbf{x})$. Ana iya kallon wannan a matsayin matsalar binciken mafi guntu hanya a cikin bishiya inda nodes suke gurbi, farashin gefe yana da alaƙa da $-\log P(x_t | gurbi)$, kuma manufar ita ce ƙididdige hanyoyi (kalmomin sirri) cikin tsari na ƙara farashin gabaɗaya (watau raguwar yuwuwar). Algorithms kamar Binciken Farashi Mai Daidaituwa (UCS) ko bambancinsa mai iyaka, Binciken Katako tare da faɗin katako mai girma da datsa mai ƙarfi, na iya cimma wannan tsari na kusan. Mahimmin abu shine cewa iyakar binciken tana da fifiko ta maki yuwuwar hanyar yanzu.

5. Sakamakon Gwaji & Bincike

5.1 Kwatance da Zato na Bazuwa

Takardar ta gabatar da sakamako masu jan hankali suna kwatanta SOPG da daidaitaccen samfurin bazuwa akan ƙirar asali iri ɗaya. Babban binciken:

• Maimaitawa Sifili: SOPG yana samar da jerin abubuwa na musamman, yayin da samfurin bazuwa ke samar da maimaitawa da yawa, yana ɓata ƙoƙarin lissafi.
• Ingantaccen Kai Hari: Don cimma irin wannan adadin rufe (kashi na kalmomin sirri a cikin gwajin gwaji da aka fasa), SOPG yana buƙatar ƙananan ƙididdiga na ƙira kuma yana samar da gabaɗaya jerin ƙanana sosai. Wannan yana fassara kai tsaye zuwa saurin fasa kalmar sirri a yanayin duniya na gaske.

5.2 Ma'auni da Mafi Kyawun Fasaha a Halin Yanzu

An yi ma'auni na SOPGesGPT da manyan ƙirar zato na kalmar sirri: OMEN (Markov), FLA, PassGAN (GAN), VAEPass (VAE), da PassGPT na zamani. A cikin gwajin wuri ɗaya:

• Adadin Rufe: SOPGesGPT ya cimma 35.06%, ya zarce OMEN da 254%, FLA da 298%, PassGAN da 421%, VAEPass da 380%, da PassGPT da 81%.
• Adadin Tasiri: Takardar kuma tana da'awar jagoranci a cikin "adadin tasiri," mai yiwuwa ma'auni ne da ke da alaƙa da inganci ko adadin bugun farko na kalmomin sirri, wanda shine babban ƙarfin SOPG.

Wannan yana nuna cewa hanyar samarwa (SOPG) tana da mahimmanci kamar tsarin ƙira don aiki.

6. Tsarin Bincike & Misalin Lamari

Tsarin Kimanta Ƙirar Zato na Kalmar Sirri:

1. Tsarin Ƙira & Horarwa: Menene cibiyar sadarwar jijiyoyi ta asali (GAN, VAE, Mai Canza Gudanar da Kai)? Yaya ake horar da shi?
2. Hanyar Samarwa: Yaya ake samar da kalmomin sirri daga ƙirar da aka horar? (misali, Zato na Bazuwa, Binciken Katako, SOPG). Wannan shine babban abin da takardar ta mayar da hankali.
3. Tsari & Ingantacciya: Shin hanyar tana samar da kalmomin sirri a cikin tsari mai amfani (raguwar yuwuwar)? Menene ingancin lissafi/zato?
4. Bambanci & Maimaitawa: Shin yana samar da sababbin kalmomin sirri ko maimaitawa da yawa?
5. Aikin Ma'auni: Adadin Rufe, Adadin Tasiri, da sauri akan daidaitattun bayanai (misali, RockYou).

Misalin Lamari Ba tare da Lamba ba: Yi la'akari da maharan biyu, Alice da Bob, suna amfani da ƙirar kalmar sirri ta GPT iri ɗaya da aka horar. Alice tana amfani da daidaitaccen samfurin bazuwa. Bob yana amfani da SOPG. Don fasa gwajin gwaji na kalmomin sirri 1000, software na Alice na iya buƙatar samar da zato miliyan 10, tare da maimaitawa 30%, don fasa 350. Software na Bob mai amfani da SOPG na iya samar da zato na musamman miliyan 1 kawai a cikin tsari mafi kyau don fasa irin wannan 350. Harin Bob yana da ingantaccen albarkatu sau 10 kuma yana kammalawa da sauri.

7. Hasashen Aikace-aikace & Hanyoyin Gaba

Aikace-aikace Nan da Nan:

• Gwajin Ƙarfin Kalmar Sirri Mai Ƙarfafawa: Ƙungiyoyin tsaro na iya amfani da ƙirar da aka haɓaka ta SOPG don bincika ƙa'idodin kalmar sirri da aka gabatar da inganci ta hanyar samar da mafi yuwuwar hanyoyin kai hari da farko.
• Maidowa Kalmar Sirri na Binciken Shari'a: Kayan aikin maidowa kalmar sirri na doka na iya haɗa SOPG don ƙara yawan nasara a cikin iyakataccen lokaci/kasafin lissafi.

Hanyoyin Bincike na Gaba:

• Ƙirar Haɗin kai: Haɗa samar da tsari na SOPG tare da ƙarfin wasu tsarin (misali, haɗa ilimin ma'ana daga manyan ƙirar harshe).
• SOPG Mai Daidaitawa/Kan layi: Canza dabarun bincike a cikin lokaci na gaske bisa ra'ayoyin daga sakamakon harin ɓangare.
• Maganganun Tsaro: Bincike cikin sabbin hanyoyin hashing na kalmar sirri ko adana fasaha waɗanda ke da juriya musamman ga harin da aka tsara, mai jagora ta yuwuwar kamar SOPG.
• Bayan Kalmomin Sirri: Yin amfani da tsarin samar da tsari zuwa wasu fannonin tsaro kamar samar da mafi yuwuwar URLs na satar bayanai ko bambance-bambancen malware.

8. Nassoshi

1. Weir, M., Aggarwal, S., Medeiros, B., & Glodek, B. (2009). Fasa Kalmar Sirri Ta Amfani da Ƙa'idodin Nahawu Mai Yuwuwar Ba tare da Mahallin Ba. A cikin IEEE Symposium on Security and Privacy.
2. Ma, J., Yang, W., Luo, M., & Li, N. (2014). Nazarin Ƙirar Kalmar Sirri Mai Yuwuwar. A cikin IEEE Symposium on Security and Privacy.
3. Hitaj, B., Gasti, P., Ateniese, G., & Perez-Cruz, F. (2017). PassGAN: Hanyar Koyo Mai Zurfi don Zato na Kalmar Sirri. A cikin International Conference on Applied Cryptography and Network Security.
4. Goodfellow, I., et al. (2014). Cibiyoyin Sadarwar Jijiyoyi Masu Gaba da Juna. Haɓaka Cibiyoyin Sadarwar Bayanai na Neuronal.
5. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Ƙirar Harshe Masu Koyon Ayyuka da yawa ba tare da Kulawa ba. OpenAI Blog.
6. Melicher, W., et al. (2016> M, Mai Ƙarfi, da Daidai: Ƙirar Yuwuwar Zato na Kalmar Sirri Ta Amfani da Cibiyoyin Sadarwar Jijiyoyi. A cikin USENIX Security Symposium.

9. Bincike na Asali & Sharhin Kwararru