Download presentation
Presentation is loading. Please wait.
Published byMillicent Bradford Modified over 8 years ago
1
Spot me if you can: Uncovering spoken phrases in encrypted VoIP conversations Charles V. Wright Scott E. Coull Gerald M. Masson Lucas Ballard Fabian Monrose Paul DiOrio Rachel Lathbury Presented by
2
The Rise of Voice Over IP Projected this year: $3.19 billion; 16.6 million subscribers This year alone: +24.3% revenue; +21.2% subscribers ● Xbox Live ● Vonage, Skype, etc ● U.S. Army Land Warrior System ● Transport for traditional telephone signals ● Many users in a chat room setting ( statistics from IBISworld ) 2 Examples:
3
SIP and (S)RTP 3 SIP: Connection set-up, connection tear down RTP: Actual transmission of audio data SRTP is increasingly used for secure RTP transmission Transports the actual voice data SRTP uses Advanced Encryption Standard in one of two cipher modes to change from a block to a stream cipher It will become clear that this encryption gives a false sense of security
4
Audio Codec Codec: program designed to encode/decode a digital signal For Audio, converts an analog signal into a digital stream ● Good for storage or transport Audio files lend themselves nicely to lossy compression ● Eliminate inaudible sounds; “easy” vs “hard” sounds 4 Generally, it searches a collection of sounds and selects the closest match
5
Speex Audio Codec Code-Excited Linear Prediction Variable Bit Rates (VBR) Encodes a window of audio samples as a frame ● Sample rates of 8kHz, 16kHz or 32 kHz ● Bit rates range from 2 – 44 kbps 5
6
VBR Encoding 6 Goal: high sound quality with less information Easier sounds to encode require fewer bits per frame VBR encoder selects the best bit rate for each frame Vowels and fricatives encode at different bit rates ∴ Packet lengths are very good indicators of what bit rate was used
7
Encrypted packets? Packet Length Bit Rate Encoding (Wright, et. al.) 7
8
8 Determining language spoken Searching for specific phrases Our research What accent is being spoken? Who is speaking? Ramifications Exploit VBR encoding and length preserving encryption
9
Some linguistic background 9 Vowels research, telephone, voice Fricatives research, telephone, voice Phoneme—smallest unit of language capable of distinguishing meaning Every language has its own native phonetic inventory and its own phonetic distribution Vowels are “harder” to encode than fricatives
10
Language Recognition (Wright, et. al.) 10
11
(Wright, et. al.) 11 Language Identification of Encrypted VoIP Traffic: Alejandro y Roberto or Alice and Bob? 2,066 speakers, 21 languages 66% accuracy (14x > random guessing) 14 languages achieve 90% Binary decisions average 86%
12
Our Research We hope to use similar techniques to discover which accent of English is being spoken We predict that individual accents will leak information despite encryption Our motivations: ● Part of a person's voiceprint ● Discover how much information can be exploited ● Save the world (or at least help) 12
13
Accents and Individuals The ultimate goal: search for an individual profile To accomplish this we will begin by making accent profiles Find linguistic differences between English accents Examine these differences and their effect on packet length Ideal: Create sufficiently dissimilar packet length distributions for each accent Likely Reality: Combination of packet length distributions and other techniques 13
14
14 (Wright, et. al.) Spot me if you can: Uncovering spoken phrases in encrypted VoIP conversations
15
Gathering the training data Used a large corpus of native English speakers We're using a corpus of non-native English speakers ● Speech Accent Archive (George Mason) ● 909 available samples Two utterances of a word will not be the same ● Intonation, rhythm, stress, etc. Use Hidden Markov Models for variation tolerance 15
16
Basic Hidden Markov Model Example: a blind hermit meteorologist using seaweed 16 model: http://www.comp.leeds.ac.uk/roger/HiddenMarkovModels/html_dev/hidden_patterns/s2_pg1.html seaweed: http://openclipart.org/people/johnny_automatic/johnny_automatic_seaweed.png weather: http://www.wyrebc.gov.uk/page.aspx?ImgID=2343http://www.comp.leeds.ac.uk/roger/HiddenMarkovModels/html_dev/hidden_patterns/s2_pg1.htmlhttp://openclipart.org/people/johnny_automatic/johnny_automatic_seaweed.pnghttp://www.wyrebc.gov.uk/page.aspx?ImgID=2343
17
Hidden Markov Model 17 (Wright, et. al.)
18
Testing techniques Limited corpus: target is “the bike” (dh ah b ay k) “the” (dh ah) “a bird” (ah b er d)[ (dh ah) (ah b) (b ay k) ] “bicameral” (b ay k ae m ax r ax l) This technique: achieved recall and precision at 0.28 ● More realistic pronunciations achieved ~0.50 Our hope: this difference shows up with accents as well 18
19
The Experiment and Results Attacker has a 1 in 3 chance of finding target phrase Results depended on specific phrases: “Young children should avoid exposure to contagious diseases” Precision: 1.0, recall:.99 “ The fog prevented them from arriving on time” Precision: 0.84, recall: 0.72 Median true positive rate was 63%, but 20% of speakers had true positive rate under 50% 19
20
Mitigation and Success 20 Languages can be recognized with as much as 90% accuracy Phrases can be located with 63% accuracy We hope accents can be found at similar rates Default SRTP encryption methods are not sufficient Padding mitigates risk Performance decrease Adding noise is not effective
21
21 Communicate at your own risk Paul DiOrio pmd8d@cs.virginia.edu Rachel Lathbury rdl5u@cs.virginia.edu Prof. Dave Evans evans@cs.virginia.edu
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.