Mining for cyber-threat intelligence to improve cyber ... · Mitigation Platform for a Trusted Internet of Things Towards threat actions Find similarity of candidate actions with

Advanced Cyber-Threat Intelligence, Detection and Mitigation Platform for a Trusted Internet of Things

Mining for cyber-threat intelligence to improve cyber-security risk mitigation

Panel on Cyber-security Intelligence2019 Community of Users Workshop

Nicholas KolokotronisDepartment of Informatics and TelecommunicationsUniversity of Peloponnese • [email protected]

mailto:[email protected]


Cyber-threat intelligence▪ From unstructured (textual)

high-volume data to

o Vulnerabilities/exploits

o Links to CVE/other VDB IDs

o Threat actors TTPs

o Specific products/platforms

o Popularity, price, …

o CVSS => measurable

▪ CTI needs to be compliant against legal requirements

2

CT


Cyber-defense goals▪ Accurate modelling of the

attack strategies

▪ Determine the attackers’ capabilities

o constraint resources (budget, tools, etc.)

▪ The attackers’ goals vary depending on the target

o access level, degrade QoS, …

▪ Define the defender’s available actions

o possible counter-measures

o highlight parameters

▪ Cyber-defenseneeds tominimizethe attacksurface

3

Dynamic risk analysis

4

Security properties should be measurable


Dynamic risk analysis: attack models

5


Example: exploitation probability▪ Need to be

measurable

o Estimated from CVSS metrics

o 𝑃 𝑒𝑖 = 2 ×𝐴𝑉 × 𝐴𝐶 × 𝐴𝑢

▪ Likewise for an attack’s attempt probability

6


ML – from CTI to structured TTPs▪ Conversion of CTIs to a semi-structured format (JSON, XML)

▪ Filtering specific (TTP, exploits) information, has the benefits:

o More easily processed in a automated way

o Only condensed information will be available

o Reports will be still readable

▪ Known formats for attack patterns is STIX v2.1

▪ The conversion of CTIs into actionable information can be achieved using ML techniques

7


Threat actions identification

8


CTI generation process

9


Classifier needed with a number of features, like:

▪ Word size (CTIs with elaborated TTPs tend to be larger)

▪ Security action word density (security correlated verbs)

▪ Security target word density (security correlated nouns)

Data pre-processing1. Need crawler that gathers all

pages from the web

o CTI vendors (e.g. Symantec)

o Forums, blogs, etc.

2. Sanitize content and keep all textual information as articles

o Remove HTML tags, images, etc.

3. Automated decision on the CTI value of each article

o otherwise it is dropped

10


[CT] CTI crawling and classification▪ Crawling components used in Cyber-Trust

11


[CT] CTI crawling and classification▪ Clear/Deep/Forum web crawling in Cyber-Trust

o Implement topic-specific crawling on publicly available web sites

▶︎ focus on Deep/Dark web sites that don’t require authentication

o Model Builder is responsible for creating the classification model; needs a set of positive and negative URLs.

o Seed Finder identifies the initial seed of URLs to crawl based on a user-defined query, e.g. on “IoT vulnerabilities”

o The crawled websites go through the Article/Forum Parser, which extracts the useful text part of each one

▶︎ internally forums are structured in a different way compared to websites

12

Dynamic risk analysis (enhanced)

13

Security properties should be measurable


Data pre-processing▪ Security correlated verbs/nouns are extracted from CVEs,

CAPEC, CWE repositories using NLP techniques

o Used on each article to find all OVS (Object, Verb, Subject) triplets; these are candidate threat actions

▪ CTI contain strings that an NLP parser may not understand, such as IoCs

o To remedy this,we temporallysubstitute thesewith RegEx, e.g.:

14


TTP specific ontology

15

▪ An ontology created by TTPs provided by ATT&CK and CAPEC repositories (MITRE)

Class name Class description Example

Kill chain phase Phase information, e.g. name or order Control or 5

Tactic Description of how to achieve a phase Privilege escalation

Technique Description of how to achieve a tactic DLL injection

Threat action Verb associated with malicious action Overwrite, Terminate

Object The action’s target File, Process

Pre-condition Action prerequisites that have to hold User access

Intent Goal/subgoal of an action Run malicious code


Towards threat actions▪ Find similarity of candidate actions with all records in ontology

▪ Information Retrieval (IR) scoring vs. threshold

▪ Vocabulary based on synonyms (e.g. by WordNet) or custom

▪ Best scoring class is assigned to the threat action

16


[CT] CTI classification▪ Topic vocabulary in Cyber-

Trust

o XML docs converted into text via XML Data Retriever

o Normalizer drops symbols, converts to lowercase, etc.

o Collected tags are multi-word terms given to Multi-Word Expression Tokenizer

▶︎“exploit kits” => “exploit-kits”

o Word2Vec finds the similarity

17


[CT] CTI classification▪ Example top terms in Cyber-Trust collection for tag ddos


CTI sharing: using STIX ▪ Structured language for

any CTI

o wide range use cases support

o can focus on relevant aspects

▪ High level of recognition by CSIRTs and LEAs

▪ Combined with TAXII 2.0

o OSS implementations

▪ Supported by MISP

Attack pattern SDO

{

“type” : “attack”,

“id” : “attack-pattern-xyz…”,

“created” : “2017-06-8T08:17:27.000Z”,

“modified” : “2017-06-8T08:17:27.000Z”,

“name” : “Input Capture”,

“description” : “Adversary logs

keystrokes to obtain credentials”,

“kill_chain_phases” : “Maintain”,

“external_references” :

[ {

“source_name” : “ATT&CK”,

“id” : “T1056”

} ]

}

19


CTI sources’ quality aspects▪ Existence of conflicting data among sources

▪ Techniques can be used to assess the credibility of source

o Using special-purpose ranking engines (e.g. SimilarWeb)

▶︎ A combination of metrics (page views, unique site users, web traffic, etc.)

▶︎ Include some Dark Web sites

o Number of users (useful for Dark Web sites)

o Number of posts per day

o Number of CVEs per day

▶︎ More than 3/4 of vulnerabilities are publicly reported online ~7d before NVD

▶︎ Mainly concerns Dark Web, paste sites, and cyber-criminal forums

20


Use of CTI in Cyber-Trust

21

CTI sharing

dark web

deep web

clear web


Conclusions - challenges▪ ML can be used for extracting CTIs to structured and

actionable formats

▪ Technical challenges for coping with heterogeneity and volume of cyber-threat data

o Need for (semi-)automated means of processing

o Focused and topic-based crawling can improve performance

o Deep/dark web exploration presents additional challenges

o Big data management and NoSQL stores for efficiency

▪ Legal compliance and privacy-preserving data mining?

22

Mining for cyber-threat intelligence to improve cyber ... · Mitigation Platform for a Trusted Internet of Things Towards threat actions Find similarity of candidate actions with

Documents