On the Semantic Patterns of Passwords and their Security Impact

On the Semantic Patterns of Passwords and their Security ImpactRAFAEL VERAS, CHRISTOPHER COLLINS, JULIE THORPE

UNIVERSITY OF ONTARIO INSTITUTE OF TECHNOLOGY

PRESENTER: KYLE WALLACE

A Familiar Scenario…

Password:

“What should I pick as my new password?”

User Name: CoolGuy90

A Familiar Scenario…

“Musical!Snowycat90”

A Familiar Scenario… But how secure is “Musical!Snowycat90” really? (18 chars)

◦ “Musical” – Dictionary word, possibly related to hobby◦ “!” – Filler character◦ “Snowy” – Dictionary word, attribute to “cat”◦ “cat” – Dictionary word, animal, possibly pet◦ “90” – Number, possibly truncated year of birth

15/18 characters are related to dictionary words!Why do we pick the passwords that we do?

Password Patterns? “Even after half a century of password use in computing, we still do not have a deep understanding of how people create their passwords” –Authors

Are there ‘meta-patterns’ or preferences that can be observed across how people choose their passwords?

Do these patterns/preferences have an impact on security?

Contributions Use NLP to segment, classify, and generalize semantic categories

Describe most common semantic patterns in RockYou database

A PCFG that captures structural, semantic, and syntactic patterns

Evaluation of security impact, comparison with previous studies

Contributions Use NLP to segment, classify, and generalize semantic categories

Describe most common semantic patterns in RockYou database

A PCFG that captures structural, semantic, and syntactic patterns

Evaluation of security impact, comparison with previous studies

Segmentation Decomposition of passwords into constituent parts

◦ Passwords contain no whitespace characters (usually)◦ Passwords contain filler characters (“gaps”) between segments

Ex: crazy2duck93^ -> {crazy, duck} & {2,93^}

Issue: What about strings that parse multiple ways?

Coverage Prefer fewer, smaller gaps to larger ones Ex: Anyonebarks98 (13 characters long)

Splitting Algorithm Source corpora: Raw word list

◦ Taken from COCA (Contemporary Corpus of American English)

Trimmed version of COCA:◦ 3 letter words: Frequency of 100+◦ 2 letter words: Top 37◦ 1 letter words: a, I

Also collected list of names, cities, surnames, months, and countries

Splitting Algorithm Reference Corpus: Collection of N-Grams, where N=3 (Full COCA)◦ N-Gram: Sequence of tokens (words)

Ex: “I love my cats”◦ Unigrams: I, love, my, cats (4)◦ Bigrams: I love, love my, my cats (3)◦ Trigrams: I love my, love my cats (2)

Common Words

Part-of-Speech Tagging Necessary step for semantic classification◦ Ex: “love” is a noun (my true love)

and a verb (I love cats)

Given segments , returns Gap segments are not tagged

Semantic Classification Assigns a semantic classifier to each password segment

◦ Only assigned to nouns and verbs

WordNet: A graph of concepts expressed as a set of synonyms◦ “Synsets” are arranged into hierarchies, more general at top

Fall back to source corpora for proper nouns◦ Tag with female name, male name, surname, country, or city

Semantic Classification

Tags represented asword.pos.#, where # is the WordNet ‘sense’

Semantic Generalization Where in the synset hierarchy should we represent a word?

Utilize a tree cut model on synset tree◦ Goal: Optimize between parameter & data description length

W=1000 (gold), W=5000 (red), W=10000(blue)

Contributions Use NLP to segment, classify, and generalize semantic categories

Describe most common semantic patterns in RockYou database

A PCFG that captures structural, semantic, and syntactic patterns

Evaluation of security impact, comparison with previous studies

Classification RockYou leak (2009) contained over 32 million passwords

Effect of generalization can be seen in a few cases (in blue)◦ Some generalizations better than

others (Ex: ‘looted’ vs ‘bravo100’)

Some synsets are not generalized (in red)◦ Ex: puppy.n.01 -> puppy.n.01

Summary of Categories Love (6,7) Places (3, 13) Sexual Terms (29, 34, 54, 69) Royalty (25, 59, 60) Profanity (40, 70, 72) Animals (33, 36, 37, 92, 96 100)

Food (61, 66, 76, 82, 93) Alcohol (39) Money (46, 74) *Some categories expanded from two letter acronyms +Some categories contain noise from names dictionary

Top 100 Semantic Categories

Contributions Use NLP to segment, classify, and generalize semantic categories

Describe most common semantic patterns in RockYou database

A PCFG that captures structural, semantic, and syntactic patterns

Evaluation of security impact, comparison with previous studies

Probabilistic Context-Free Grammar A CFG whose productions have associated probabilities

◦ A vocabulary set (terminals) ◦ A variable set (non-terminals) ◦ A start variable ◦ A set of rules (terminals + non-terminals)◦ A set of probabilities on rules, such that

Semantic PCFG In the author’s PCFG:

◦ is comprised of the source corpora and learned gap segments◦ is the set of all semantic and syntactic categories◦ All rules are of the form , or (nonterminals)

This grammar is regular (described by a finite automaton)

Sample PCFG Training data:

◦ iloveyou2◦ ihatedthem3◦ football3

rules are base structures

Grammar can generate passwords

Probability of a password is the product of all rule probabilities

Ex: P(youlovethem2) = 0.0103125

RockYou Base Structures (Top 50)

Contributions Use NLP to segment, classify, and generalize semantic categories

Describe most common semantic patterns in RockYou database

A PCFG that captures structural, semantic, and syntactic patterns

Evaluation of security impact, comparison with previous studies

Building a Guess Generator Cracking attacks consist of three steps:

◦ Generate a guess◦ Hash the guess using the same algorithm as target◦ Check for matches in the target database

Most popular methods (using John the Ripper program)◦ Word lists (from previous breaks)◦ Brute force (usually after exhausting word lists)

Guess Generator

At a high level:◦ Output terminals in highest

probability order◦ Iteratively replaces higher

probability terminals with lower probability ones

◦ Uses priority queue to maintain order

Will this produce the same list of guesses every time?

Guess Generator Example Suppose only one base structure:

Initialized with most probable terminals: “I love Susie’s cat” Pop first guess off queue (“IloveSusiescat”)

◦ Replace first segment: “youloveSusiescat”◦ Replace second segment: “IhateSusiescat”◦ Replace third segment: “IloveBobscat”◦ Replace fourth segment: “IloveSusiesdog”

Mangling Rules Passwords aren’t always strictly lowercase◦ Beardog123lol ◦ bearDOG123LoL ◦ BearDog123LoL

Three types of rules:◦ Capitalize first word segment◦ Capitalize whole word segment◦ CamelCase on all segments

Any others?

Comparison to Weir Approach Author’s approach seen as an evolution of Weir

◦ Weir contains far fewer non-terminals (less precise estimates)◦ Weir does not learn semantic rules (fewer overall terminals)◦ Weir treats grammar and dictionary input separately

◦ Authors semantic classification needs to be re-run for changes

Password Cracking Experiments Considered 5 methods:

◦ Semantic approach w/o mangling rules◦ Semantic approach w/ custom mangling rules◦ Semantic approach w/ JtR’s mangling rules◦ Weir approach◦ Wordlist w/ JtR’s default rules + incremental brute force

Attempted to crack LinkedIn and MySpace leaks

Experiment 1: RockYou vs LinkedIn

5,787,239 unique passwords Results:

◦ Semantic outperforms non-semantic versions

◦ Weir approach is worst (67% improvement)

◦ Authors approach is more robust against differing demographics

Experiment 2: RockYou vs MySpace

41,543 unique passwords Results:

◦ Semantic approach outperforms all◦ No-rules performs best

◦ Weir approach is worst (32% improvement)

◦ Password were phished, quality lowered?

Experiment 3: Maximum Crack Rate

Since method is based on grammar, can build grammar recognizer to check Results:

◦ Semantic equivalent to brute force, with fewer guesses

◦ Weir approach generates fewer guesses, 30% less guessed

Experiment 3: Time to Maximum Crack

Fit non-linear regression to sample of guess probs. Results:

◦ Semantic method has lower guess/second

◦ Grammar is much larger than Weir method

Issues with Semantic Approach Further study needed into performance bottlenecks

◦ Though semantic method is more efficient (high guesses/hit)

Approach requires a significant amount of memory◦ Workaround involves probability threshold for adding to queue

Duplicates could be produced due to ambiguous splits◦ Ex: (one, go) vs (on, ego)

Conclusions There are underlying semantic patterns in password creation

These semantics can be captured in a probabilistic grammar

This grammar can be used to efficiently generate probable passwords

This generator shows (up to) a 67% improvement over previous efforts

Thank you!QUESTIONS?