IoT Malware Detection through Threshold Random Walks
2 likes972 views
This document presents a master's thesis on detecting IoT malware using a threshold random walk algorithm, addressing the vulnerabilities of IoT devices due to low security. It discusses the design and implementation of a lightweight, efficient detection system capable of identifying attacks while utilizing limited computing resources. The thesis also covers related works, a distributed honeypot architecture for capturing attack samples, and experimental results that demonstrate the algorithm's promising performance in detecting unknown attack patterns.
IoT Malware Detection through Threshold Random Walks
- 1. Candidate: Botticelli Biagio Advisor: Prof. Leonardo Querzoni Co-Advisor: Dott. Giuseppe Laurenza Master of Science in Engineering in Computer Science - A.Y. 2016 - 2017 IoT Malware Detection through Threshold Random Walks
- 2. Anonymity in the Web based on Routing ProtocolsIoT Malware detection through Threshold Random Walks 2 Internet of Things The Internet of Things describes the vision where objects become part of the Internet: where every object is uniquely identified, and accessible to the network, its position and status known, where services and intelligence are added to this expanded Internet, fusing the digital and physical world into a single one. 6.5 Devices per Person An increased connectivity leads to an exponential increase in the threat surface: more smart technology we add, more likely is to be hacked from the point of view of security.
- 3. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 3 Mirai Example: IoT as a weapon 20th September 2016 : KrebsOnSecurity.com targeted by an extremely large and unusual Distributed Denial-of-Service (DDoS) attack of over 660 Gbps of traffic. Innovative Aspect: the attack was performed by using direct traffic generated by a botnet of hacked IoT devices infected by a malware called Mirai. ThingsBot (or Botnet of Things): automated botnet of compromised IoT devices (things). Botmaster Botnet: robot network of hacked machines (or bots), which run malicious code under the remote command and control (C&C) of a botmaster for many malicious activities. IoT as weapon: from Internet of Things to Internet of Threats!
- 4. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 3 Mirai Example: IoT as a weapon 20th September 2016 : KrebsOnSecurity.com targeted by an extremely large and unusual Distributed Denial-of-Service (DDoS) attack of over 660 Gbps of traffic. Innovative Aspect: the attack was performed by using direct traffic generated by a botnet of hacked IoT devices infected by a malware called Mirai. ThingsBot (or Botnet of Things): automated botnet of compromised IoT devices (things). Botmaster Botnet: robot network of hacked machines (or bots), which run malicious code under the remote command and control (C&C) of a botmaster for many malicious activities. IoT as weapon: from Internet of Things to Internet of Threats!
- 5. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 4 IoT Malware AidraMirai Tsunami Linux.Hydra Chuck Norris Psyb0t Hajime Linux/IRCTelnet LightAidra RemaintenBASHLITE Predecessor Successor Influenced LEGEND 2008 200920102010 2013 2014 2016 2016 2016 2016 2014
- 6. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 5 Related Works • Honeypharm: “the more honeypots there are in different networks , the higher are the chances to capture new malware samples”. Key Concept: Distributed Architecture of low-interaction honeypots • IoTPOT: “the more embedded services are emulated for different CPU architectures, the more information on existing IoT malware can be obtained”. Key Concept: Emulation of embedded services for different architectures • SIPHON: “rather than emulated embedded services, the use of real-existing high interactive vulnerable IoT devices improves results in attracting cyber-criminals”. Key Concept: Real embedded vulnerable IoT devices offered to attackers • Fast Port-scan Detection using SHT: ”the Threshold Random Walk algorithm could be used to effectively detect the reconnaissance phase of network attacks”. Key Concept: Threshold Random Walk applied for Malware Detection
- 7. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 6 Problem Statement & Thesis Contributions Problem Statement: devices of the Internet of Things are under constant attack of cyber-criminals since they are typically low secured (or completely unsecure). However, we cannot adopt traditional lines of defense for malware detection due to computational resource constraints. Thesis Contributions: design and implement an online detection Threshold Random Walk- based algorithm which is fast, light and capable to identify attacks even with the low resources of Internet of Things sensors and objects. To get more knowledge of attack techniques performed by IoT malware, a Distributed Architecture of honeypots had been implemented. This architecture should attract modern attack patterns and capture samples of the newest threats from different locations in the world.
- 8. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 7 Distributed Honeypot Architecture Automated Procedure: the DIAG VM daily connects to Cowrie instance in New York and to Cowrie-Dumper in Singapore to locally download all the obtained data (logs and malware samples) and to restore the initial honeypot configuration. Cowrie in New York IP: 162.243.211.8 Cowrie-Dumper in Singapore IP: 128.199.204.0 DIAG VM in Rome IP: 192.168.2.197 DIAG Network Results: a total number of 332 970 attacking sessions were collected (~100 Gb of data).
- 9. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 8 Distribution of Top 15 Attacking IPs - NY New York Cowrie: 294 943 connections, 53 718 originated by different IPs.
- 10. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 9 Distribution of Top 15 Attack IPs - Singapore Singapore Cowrie-Dumper: 50 897 connections, 15 250 originated by different IPs. Observation 2: Only 299 IPs attacked both New York and Singapore honeypot instances. Observation 1: Italy is 18th with 133 IPs.
- 11. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 10 Threshold Random Walk η1 η0 η2 time Y1 Y2 Y3 Y4 Y5 Y6 Y7 Y8 Y9 Y10 Λ(Y) WARNING H1 = ATTACK H0 = LEGAL
- 12. Anonymity in the Web based on Routing ProtocolsIoT Malware detection through Threshold Random Walks 11 Attack Patterns & Attack Groups • Initial Shell Pattern • Busybox & Busybox Checks • Connectors • Malware Download • Hexadecimal Code • Malware Creation • System Exploration • Kill Processes • Fingerprinting • Suspect Files Dangerousness: the degree of danger of the command is given by the command type, contextualized within the type of interaction that we are considering. Attack Groups: “fast” data structures in which to store attack strings, classified according to their dangerousness.
- 13. Anonymity in the Web based on Routing ProtocolsIoT Malware detection through Threshold Random Walks 11 Attack Patterns & Attack Groups • Initial Shell Pattern • Busybox & Busybox Checks • Connectors • Malware Download • Hexadecimal Code • Malware Creation • System Exploration • Kill Processes • Fingerprinting • Suspect Files Dangerousness: the degree of danger of the command is given by the command type, contextualized within the type of interaction that we are considering. Attack Groups: “fast” data structures in which to store attack strings, classified according to their dangerousness. Dangerous Attack Probability: 99% Knowledge Base
- 14. Anonymity in the Web based on Routing ProtocolsIoT Malware detection through Threshold Random Walks 11 Attack Patterns & Attack Groups • Initial Shell Pattern • Busybox & Busybox Checks • Connectors • Malware Download • Hexadecimal Code • Malware Creation • System Exploration • Kill Processes • Fingerprinting • Suspect Files Dangerousness: the degree of danger of the command is given by the command type, contextualized within the type of interaction that we are considering. Attack Groups: “fast” data structures in which to store attack strings, classified according to their dangerousness. High Attack Probability: 90% Dangerous Attack Probability: 99% Knowledge Base
- 15. Anonymity in the Web based on Routing ProtocolsIoT Malware detection through Threshold Random Walks 11 Attack Patterns & Attack Groups • Initial Shell Pattern • Busybox & Busybox Checks • Connectors • Malware Download • Hexadecimal Code • Malware Creation • System Exploration • Kill Processes • Fingerprinting • Suspect Files Dangerousness: the degree of danger of the command is given by the command type, contextualized within the type of interaction that we are considering. Attack Groups: “fast” data structures in which to store attack strings, classified according to their dangerousness. High Attack Probability: 90% Medium Attack Probability: 70% Dangerous Attack Probability: 99% Knowledge Base
- 16. Anonymity in the Web based on Routing ProtocolsIoT Malware detection through Threshold Random Walks 11 Attack Patterns & Attack Groups • Initial Shell Pattern • Busybox & Busybox Checks • Connectors • Malware Download • Hexadecimal Code • Malware Creation • System Exploration • Kill Processes • Fingerprinting • Suspect Files Dangerousness: the degree of danger of the command is given by the command type, contextualized within the type of interaction that we are considering. Attack Groups: “fast” data structures in which to store attack strings, classified according to their dangerousness. High Attack Probability: 90% Medium Attack Probability: 70% Low Attack Probability: 60% Dangerous Attack Probability: 99% Knowledge Base
- 17. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 12 TRW as Binary Classification Problem TRW detection is a binary classification problem in which the output is chosen among two hypotheses: • TP - Detection: TRW selects H1, detecting the interaction as an attack and H1 is in fact True. • FP - False Positive - Type I Error: TRW selects H1 (attack) when H0 is in fact True; TRW receives a legitimate interaction as input and it detects the connection as malicious. • FN - False Negative - Type II Error: TRW chooses H0 (legal), but H1 was True; TRW receives a malicious interaction as input and it detects the connection as legitimate. • TN - Nominal: TRW picks H0 when H0 is in fact True. Binary Classification Confusion Matrix
- 18. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 13 How does the Threshold Random Walk perform? Is it correctly formulated? Experiment 1: k-Fold Cross Validation Dataset 1: all attacking sessions captured by honeypots between 24th April and 31st October 2017. 270 379 malicious interactions in total. k-Fold Cross Validation: the data is divided into k subsets of the same size. Each one of the k subsets is used once as the validation set and the other k−1 subsets are put together to form the training set. In cases of large imbalance in the dataset, stratified approach folds are created containing approximately the same percentage of samples of each target class as the complete set.
- 19. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 14 Experiment 1: Average Metrics Results Standard Deviation
- 20. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 15 TRW has very good performances even on potentially unknown attack sessions formed by new attack strings never seen before. Experiment 2: Metrics Results Dataset 2: all “new” attacking sessions captured in the last months of November and December 2017. 125 182 total interactions: equally divided in 62 591 new malicious and 62 591 legal logs. How does the Threshold Random Walk perform in case of “unknown” attacking sessions? How does the algorithm behave in terms of number of commands necessary to carry out the detection?
- 21. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 16 Experiment 2: Detection Performances Threshold Random W a l k d e t e c t s a malicious series of commands in ~ 6,44 events on average with a maximum of 9 commands required.
- 22. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 17 Experiment 3: Detection vs. Execution Average Length at Detection Dataset 3: all “complete” attacking sessions formed by series of commands that would actually infect a device. 114 226 logs = ~34.305% of 332 970 total interactions Each interaction has the characteristic of having at least one command to sample execution.
- 23. IoT Honeypots: State of the ArtIoT Malware detection through Threshold Random Walks 18 Conclusions & Future Works Conclusions: experiments proved that Threshold Random Walk outcomes promising results. It’s: • Fast: detection of malicious interactions id performed in early stages of attacking sessions; • Lightweight: no particular computing requirements; • Extensible: upgradeable knowledge base allows to include emerging new attack techniques; • Efficient: TRW makes the correct decision on quite all observed malicious sequences of commands. Future Works: • SSH/Telnet Emulation: create a Telnet/SSH traffic sampling system on a real communication channel. • Automation of KB Creation: design an automated process that integrates into the existing KB new discovered attack strings, without necessarily having to start its creation from scratch. • Architecture Improvement: new honeypot solution could be integrated in the existing architecture. • ELK Framework: ElasticSearch, LogStash and Kibana Data Analytics tool could be integrated into the DIAG VM server to have a visual report of collected data in structured file formats (.json files).
- 24. “A secure system is one that does what is supposed to do, and nothing more”. J.B. Ippolito, Native Intelligence, Inc. Any Question?
- 25. Biagio Botticelli - botticelli.1212666@studenti.uniroma1.it M.Sc. in Engineering in Computer Science Thank You! “A secure system is one that does what is supposed to do, and nothing more”. J.B. Ippolito, Native Intelligence, Inc. Any Question?