1 / 48

Paola García-Perera, Carlos Mex-Perera and Juan A. Nolazco-Flores

A phoneme-space-representation heuristic to improve the performance in a cryptographic-speech-key generation task. Paola García-Perera, Carlos Mex-Perera and Juan A. Nolazco-Flores Dpto. De Ciencias Computacionales ITESM, campus Monterrey jnolazco@itesm.mx. Talk's Overview. Introduction

todd
Download Presentation

Paola García-Perera, Carlos Mex-Perera and Juan A. Nolazco-Flores

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A phoneme-space-representation heuristic to improve the performance in a cryptographic-speech-key generation task Paola García-Perera, Carlos Mex-Perera and Juan A. Nolazco-Flores Dpto. De Ciencias Computacionales ITESM, campus Monterrey jnolazco@itesm.mx

  2. Talk's Overview • Introduction • Proposal • Scenarios • Challenges • Solutions • Experiments and Results • Comments and Conclusions

  3. Introduction • In the last years Computer Security is growing in importance, this is specially motivated for the following reasons: • Estimated worldwide loss in 2002 resulting from security breaches is around 1.6 TRILLIONS, USD. • The increase of highly sophisticated tools available in Internet and the easy of use of this tools!!!

  4. High Tools packet spoofing denial of service Wireless Intruder Knowledge sniffers sweepers automated probes/scans GUI back doors network mgmt. diagnostics disabling audits hijacking sessions burglaries exploiting known vulnerabilities Attack Sophistication password cracking self-replicating code Attackers password guessing Low 1980 1985 1990 1995 2000 2005 Source: Software Engineering Institute, CMU, 2000 Introduction: Tool sophistication & Intruder Knowledge “stealth” / advanced scanning techniques DDOS attacks www attacks

  5. Introduction • SINGAPORE Nov 13, 2003– Two 15-year-old hackers have been placed under probation for two years, one for flooding his teacher’s e-mail inbox with more than 160,000 messages... • Michael Buen and Onel de Guzman, the 20-something Filipino college students who allegedly wrote the ILOVEYOU virus.

  6. Introduction • So, it is urgent to develop secure systems!! • Secure systems features: • Confidenciality (Privacity) • Integrity • Availablity • However, it is not simple to make sure a system have all these features, therefore technologies of different kinds should be used.

  7. Introduction

  8. Introduction • In this work we will be interested in • Biometrics and • Criptography

  9. Introduction

  10. Introduction • Depending on the configuration of the Cryptographic systems, it will allow a Secure System to have: • Confidenciality (C) also called Privacity and/or • Integrity (I), and/or • Availablity (not in this work)

  11. Introduction • Authenticity needs the following factors: • Possesion: Encripted Keyfile i.e. in a Smart Card • Knowledge: Password or Identity Number; or Biometric: something you are (fingerprint, iris, etc.)

  12. Intruder User Password Message Channel or Encryption Decryption Message Source data storage Key Key 1. A, S(C,I) Easy to Decipher, Low Computational Complexity The Password and Key can be stolen allowing Ai,S(~C,~I) and DoS Keys Generator Intoduction:Cryptosystem

  13. Introduction:Key's features The key defines exactly the transformations that take place in the encryption and decryption stages. The key should be very difficult to guess and very difficult to reproduce.

  14. Intruder User Password Message Channel or Encryption Decryption Message Source data storage KeyPriv KeyPub 2. A,S(C, I); A thanks to the password Very Difficult to Decipher, High Complexity The Password and Key can be stolen allowing Ai,S(C, I), DoS Keys Generator Intoduction:Cryptosystem

  15. Intruder User Password Message Channel or Encryption Decryption Message Source data storage KeyPub KeyPriv 3. A,S(C, I); S(C,I) thanks to the password Very Difficult to Decipher, High Complexity The Password and Key can be stolen allowing Ai,S(C, I), DoS Keys Generator Intoduction:Cryptosystem

  16. IntroductionRequirment Generate a strong cryptographic key.

  17. IntroductionObservations • A strong cryptographic keycan be obtained using biometrics because once it is used, the key can be destroyed!!! • Moreover, the biometric can also be used as password!! • Using speech allows flexibility to generate a large number of sentences, given the posibility to easily change the passphrase.

  18. Proposal Generate a strong cryptographic key using speech.

  19. Proposal:Scenarios

  20. Proposal:Another Applications • Reading encrypted mail through telephone. • Secure conversation through telephone

  21. Proposal:Technologies Needed • Speech Tecnology • Pattern Matching • Cryptographic concepts

  22. Utterance's phonemes recognition and segments alignment utterance Phoneme's characterization computation Key Key generation Proposal:Proposed solution's block diagrams

  23. Proposal:example Uttered Sentence: This is my key This is my key /DH/ /IH/ /S/ /IH/ /Z/ /M/ /AY/ /K/ /IY/ spkr1: 0 1 0 0 0 1 0 1 1 spkr2: 1 1 0 1 1 0 0 1 0 spkr3: 1 0 1 1 0 0 1 0 1

  24. Proposal:Challenges • Accurate utterance's phoneme recognition and segmentations. • Phoneme characterization. • Generate the same key every time the same user utters a sentence, and this key should be different to the key generated for other users uttering the same sentence. • Control of the decision hyperplane.

  25. Proposed Solutions • The system asks the user to utter a sentence. • User utters the sentence. • Using a force-aligment ASR both the sentence's phonemes and their starting and ending in the speech utterance are obtained. • Phonemes are characterised. • A SVM is used to generate the key.

  26. sentence phoneme's HMM model Force-aligment ASR

  27. /p/ /p/ /p/ /p/ M( R ) /p/ /p/ /p/ M( R ) M( R ) M( R ) M( R ) M( R ) i,j i,j i,j i,j i,j i,j /f/ HMM /n/ HMM /n/ HMM /o/ HMM + + + + + + + + 1 1 0 1 1 0 0 SVM Heuristic SVM Heuristic SVM Heuristic SVM Heuristic SVM Heuristic SVM Heuristic SVM Heuristic

  28. Block Diagram

  29. Other issues • Size of the feature space: • We know that for the ASR task a proper number of parameters is 13 MFCC. However, that may not be the case when using the SVM (as we can see in the next slides). • Phoneme-space-representation heuristic • Models for 10, 12, 14, 16, 18, 20 and 22 MFCC were trained. Then the models with highest accuracy for each phoneme were selected. • As a result, the accuracy using this fusion outperformance to individual model.

  30. SVM Heuristic • For instance the values for η for 10 users of the phoneme /AH/ are: • Then , we choose model 20 to be the final model. • The same method is employed for different number of users and phonemes.

  31. Experiments and Results • We use YOHO database (LCD database) • 138 speakers • three pair of numbers are uttered. • 8 Gaussian Continuous HMM were used for phonemes's acoustic models.

  32. Experiments and Results • SVM with RBF kernel and choosing the best model. • The effect of the selection of a specific number of parameters is shown in the following graph.

  33. η for different number of parameters

  34. Global average for η, after the best performance models 1 Previous work using Gaussian weights (no SVM tuning) 2 RBF tuning parameters for better solution (no Gaussian weight consideration) 3 RBF best first results (SVM raw parameters)

  35. Comments and Conclusions • An arquitecture to obtain a cryptographic-speech key. • The results shown feasibility of the method to obtain cryptographic-speech keys, performing a simple heuristic. • The higher computational load is performed offline. • Future work: • Try noise environments. • M-ary classification • Try error correction concepts

  36. Comments and conclusiones • Related Work, Bell Labs • Text-Independent • F. Monrose, M. K. Reiter, Q. Li and S. Wetzel. Cryptographic key generation from voice. In Proceedings of the 2001 IEEE Symposium on Security and Privacy, May 2001.

  37. Comments and conclusiones • Related Work (CMU)

  38. Comments and conclusiones • Related Work (CMU)

  39. FIN

  40. Arrivederchi!!!

  41. Graphic Representation

  42. Number of Users η ACA(%) 10 88.54 20 85.77 30 84.24 50 81.89 Table1. Value of η for different number of users Table of Results

  43. CHALLENGES

  44. Proposal:Our Solutions • Accurate utterance's phoneme recognition and segmentations. • Make the system to know the sentence (generated randomly to avoid the intruder to record the voice). • Apply force-aligment ASR to obtain the segments of the phonemes.

  45. Proposal:Our Solutions • Phoneme characterization. • Train acoustic models of the phonemes for each user. • The phonemes segments • Once we know the phonemes and its segments, we characterise the phonemes in the utterance by using MFCC from • the phoneme segment and • from the acoustic model of the phoneme.

  46. Proposal:Our Solutions • Generate the same key every time the same user utters a sentence, and this key should be different to the key generated for other users uttering the same sentence. • Given the phoneme's characterization we used a classifier: SVM (Support vector machine), to obtain a bit for each phoneme of the utterance. • The stream of bits will be the key.

  47. Support Vector Machine The goal of basic SVM is to obtain a model to perform vector classification in one of two classes.

  48. Support Vector MachineNon-separable case • For a better classification the input data is mapped into a higher dimensional space by function . • The exact specification of  is not necessary; instead, a kernel is defined.

More Related