Koyon Injiniya na Adawa don Ƙimar Ƙarfin Kalmar Sirri mai Ƙarfi

Teburin Abubuwan Ciki

• 1. Gabatarwa
• 2. Tarihi da Ayyuka masu Alaka
• 3. Hanyoyin Aiki
• 4. Saitin Gwaji
• 5. Sakamako da Tattaunawa
• 6. Cikakkun Bayanai na Fasaha da Tsarin Lissafi
• 7. Misalin Tsarin Bincike
• 8. Aikace-aikace da Hanyoyi na gaba
• 9. Bincike na Asali
• 10. Manazarta

1. Gabatarwa

Kalmomin sirri sun kasance ginshiƙin tsaro na dijital, duk da haka zaɓin kalmar sirri mara ƙarfi yana fallasa masu amfani ga manyan haɗari. Masu ƙididdige ƙarfin kalmar sirri na gargajiya sun dogara da ƙa'idodin lexical na tsaye (misali, tsayi, bambancin haruffa) kuma sun kasa daidaitawa da hare-haren adawa masu tasowa. Wannan binciken yana ba da shawarar koyon injiniya na adawa (AML) don horar da samfura akan kalmomin sirri na yaudara da aka ƙera da gangan, don inganta ƙarfi. Ta amfani da tarin bayanai sama da samfuran kalmar sirri na adawa 670,000 da algorithms na rarrabuwa guda biyar, marubutan sun nuna har zuwa 20% ingantuwa a daidaiton rarrabuwa akan samfuran gargajiya.

2. Tarihi da Ayyuka masu Alaka

Kayan aikin da ake da su kamar Password Meter, Microsoft Password Checker, da Google Password Meter suna amfani da ƙa'idodin tsaye. Duk da haka, kalmomin sirri na adawa—kamar 'p@ssword' maimakon 'password'—suna amfani da waɗannan ƙa'idodin, suna haifar da kuskuren rarrabuwa. Hare-haren adawa a cikin koyon injiniya, kamar yadda Goodfellow et al. (2014) suka yi nazari, sun haɗa da ƙirƙirar abubuwan shigar da ke yaudarar samfura. Wannan aikin ya faɗaɗa wannan ra'ayi zuwa ƙididdige ƙarfin kalmar sirri, wani yanki da ba a bincika sosai ba.

3. Hanyoyin Aiki

Marubutan sun yi amfani da algorithms na rarrabuwa guda biyar: Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), da Neural Network. Tarin bayanai ya ƙunshi sama da samfuran kalmar sirri na adawa 670,000, kowanne an lakafta shi a matsayin mai rauni, matsakaici, ko mai ƙarfi. Horar da adawa ya haɗa da ƙara tarin bayanan horarwa tare da misalan adawa da aka samar ta hanyoyi kamar Fast Gradient Sign Method (FGSM) da Projected Gradient Descent (PGD).

4. Saitin Gwaji

An gudanar da gwaje-gwaje akan tsarin bututun koyon injiniya na yau da kullun tare da rabon horo-gwaji na 80-20. Ma'aunin kimantawa sun haɗa da daidaito, daidaiton hasashen, tunawa, da makin F1. An horar da samfuran tushe akan bayanai masu tsabta, yayin da aka horar da samfuran adawa akan bayanai da aka ƙara waɗanda suka haɗa da misalan adawa.

5. Sakamako da Tattaunawa

Horar da adawa ya inganta daidaito har zuwa 20% a duk masu rarrabuwa. Misali, daidaiton Random Forest ya ƙaru daga 72% zuwa 86%, kuma Neural Network daga 75% zuwa 90. Matrix rikicewa ya nuna raguwa mai mahimmanci a cikin ƙididdigar ƙarya (kalmomin sirri marasa ƙarfi da aka rarraba a matsayin masu ƙarfi). Binciken ya nuna cewa horar da adawa ba wai kawai yana karewa daga hare-haren da aka sani ba, har ma yana faɗaɗa zuwa tsarin adawa da ba a gani ba.

6. Cikakkun Bayanai na Fasaha da Tsarin Lissafi

Manufar horar da adawa za a iya tsara ta azaman rage mafi munin asara akan ɓarna na adawa:

$\min_{\theta} \mathbb{E}_{(x,y) \sim \mathcal{D}} \left[ \max_{\delta \in \mathcal{S}} \mathcal{L}(f_\theta(x+\delta), y) \right]$

inda $\theta$ sune sigogin samfur, $\mathcal{D}$ shine rarraba bayanai, $\delta$ shine ɓarnar adawa da aka ƙuntata zuwa saitin $\mathcal{S}$ (misali, $\|\delta\|_\infty \leq \epsilon$), kuma $\mathcal{L}$ shine aikin asara. Ga bayanan kalmar sirri, ɓarna sun haɗa da maye gurbin haruffa (misali, 'a' zuwa '@') da saka.

FGSM yana samar da misalan adawa kamar:

$x_{adv} = x + \epsilon \cdot \text{sign}(\nabla_x \mathcal{L}(f_\theta(x), y))$

Wannan hanya tana tabbatar da cewa samfura suna koyon yin tsayayya da ƙananan ɓarna na mugunta.

7. Misalin Tsarin Bincike

Yi la'akari da kalmar sirri 'Password123'. Mai duba gargajiya zai iya rarraba ta a matsayin mai ƙarfi saboda haɗaɗɗun manya da ƙananan haruffa da lambobi. Duk da haka, bambancin adawa 'P@ssword123' (maye gurbin 'a' da '@') za a iya kuskuren rarraba shi. Tsarin da aka gabatar yana horar da samfura don gane irin waɗannan maye gurbin a matsayin marasa ƙarfi. Misalin dabaru na yanke shawara:

Shigar: kalmar sirri = "P@ssword123"
1. Duba bambancin haruffa: haɗaɗɗun manya da ƙananan haruffa, lambobi, haruffa na musamman -> makin farko: 8/10
2. Gano tsarin adawa: an gano '@' a maimakon 'a' -> hukunci: -3
3. Makin ƙarshe: 5/10 -> Mai rauni

Wannan misalin tushen ƙa'ida yana nuna halin da aka koya na samfuran adawa.

8. Aikace-aikace da Hanyoyi na gaba

Hanyar za a iya faɗaɗa zuwa wasu fannonin tsaro kamar gano baƙon saƙo, tsarin gano kutse, da tantancewar halittu. Ayyuka na gaba sun haɗa da bincika cibiyoyin sadarwa na adawa (GANs) don ƙirƙirar kalmomin sirri na adawa daban-daban, da haɗa gano adawa na ainihi cikin masu sarrafa kalmar sirri. Bugu da ƙari, koyon canja wuri zai iya ba da damar ƙarfi a fannoni daban-daban.

9. Bincike na Asali

Mahimman Bayani: Wannan takarda ta nuna gamsasshiyar hujja cewa koyon injiniya na adawa ba kawai sha'awar ka'ida ba ce amma larura ce mai amfani ga ƙididdige ƙarfin kalmar sirri. Ƙimar daidaiton 20% tana da mahimmanci, musamman a wani yanki inda ko da kuskuren rarrabuwa ɗaya zai iya haifar da ɓarnar bayanai.

Tsarin Tunani: Marubutan sun fara da gano yanayin tsaye na kayan aikin yanzu, sannan suka gabatar da misalan adawa a matsayin barazana, kuma suka ba da shawarar horar da adawa a matsayin mafita. Tabbatar da gwaji ya kasance cikakke, yana rufe masu rarrabuwa da ma'auni da yawa.

Ƙarfi da Rashi: Babban ƙarfi shine babban tarin bayanai (samfura 670k) da ingantacciyar ingantuwa a duk samfura. Duk da haka, takardar ba ta bincika farashin lissafi na horar da adawa ba, kuma ba ta gwada waɗanda ke kai hari masu daidaitawa waɗanda suka san tsaron ba. Bugu da ƙari, hanyoyin samar da adawa (FGSM, PGD) suna da sauƙi; hare-hare masu rikitarwa kamar Carlini-Wagner na iya zama mafi ƙalubale.

Bayani masu Amfani: Ga masu aiki, haɗa horar da adawa cikin masu duba ƙarfin kalmar sirri abu ne mai sauƙi. Ƙungiyoyi yakamata su sabunta manufofin kalmar sirri don haɗa masu ƙididdige tushen ML. Bincike na gaba yakamata ya mayar da hankali kan gano adawa na ainihi da ƙarfi ga hare-haren adawa masu daidaitawa. Kamar yadda Goodfellow et al. (2014) suka lura a cikin takardarsu ta asali akan misalan adawa, tseren makamai tsakanin mahara da masu tsaro yana ci gaba, kuma wannan aikin mataki ne na gaba.

10. Manazarta

1. Goodfellow, I. J., Shlens, J., & Szegedy, C. (2014). Explaining and Harnessing Adversarial Examples. arXiv:1412.6572.
2. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., & Vladu, A. (2017). Towards Deep Learning Models Resistant to Adversarial Attacks. arXiv:1706.06083.
3. Password Meter. (n.d.). Retrieved from https://www.passwordmeter.com/
4. Microsoft Password Checker. (n.d.). Retrieved from https://account.microsoft.com/security/password
5. Carlini, N., & Wagner, D. (2017). Towards Evaluating the Robustness of Neural Networks. IEEE Symposium on Security and Privacy.