kth.sePublications
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Devising and Detecting Phishing Emails Using Large Language Models
KTH, School of Electrical Engineering and Computer Science (EECS), Computer Science, Network and Systems Engineering. Harvard Univ, Harvard John A Paulson Sch Engn & Appl Sci, Cambridge, MA 02138 USA.ORCID iD: 0000-0001-7884-966x
Harvard Univ, Harvard Kennedy Sch, Cambridge, MA 02138 USA..
Avant Res Grp, Buffalo, NY 14214 USA..
MIT, Cambridge, MA 02139 USA..
Show others and affiliations
2024 (English)In: IEEE Access, E-ISSN 2169-3536, Vol. 12, p. 42131-42146Article in journal (Refereed) Published
Abstract [en]

AI programs, built using large language models, make it possible to automatically create phishing emails based on a few data points about a user. The V-Triad is a set of rules for manually designing phishing emails to exploit our cognitive heuristics and biases. In this study, we compare the performance of phishing emails created automatically by GPT-4 and manually using the V-Triad. We also combine GPT-4 with the V-Triad to assess their combined potential. A fourth group, exposed to generic phishing emails, was our control group. We use a red teaming approach by simulating attackers and emailing 112 participants recruited for the study. The control group emails received a click-through rate between 19-28%, the GPT-generated emails 30-44%, emails generated by the V-Triad 69-79%, and emails generated by GPT and the V-Triad 43-81%. Each participant was asked to explain why they pressed or did not press a link in the email. These answers often contradict each other, highlighting the importance of personal differences. Next, we used four popular large language models (GPT, Claude, PaLM, and LLaMA) to detect the intention of phishing emails and compare the results to human detection. The language models demonstrated a strong ability to detect malicious intent, even in non-obvious phishing emails. They sometimes surpassed human detection, although often being slightly less accurate than humans. Finally, we analyze of the economic aspects of AI-enabled phishing attacks, showing how large language models increase the incentives of phishing and spear phishing by reducing their costs.

Place, publisher, year, edition, pages
Institute of Electrical and Electronics Engineers (IEEE) , 2024. Vol. 12, p. 42131-42146
Keywords [en]
Phishing, large language models, social engineering, artificial intelligence
National Category
Computer Sciences
Identifiers
URN: urn:nbn:se:kth:diva-345143DOI: 10.1109/ACCESS.2024.3375882ISI: 001192203500001Scopus ID: 2-s2.0-85187996490OAI: oai:DiVA.org:kth-345143DiVA, id: diva2:1849622
Note

QC 20240408

Available from: 2024-04-08 Created: 2024-04-08 Last updated: 2024-09-18Bibliographically approved
In thesis
1. Mitigating AI-Enabled Cyber Attacks on Hardware, Software, and System Users
Open this publication in new window or tab >>Mitigating AI-Enabled Cyber Attacks on Hardware, Software, and System Users
2024 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

This doctoral thesis addresses the rapidly evolving landscape of computer security threats posed by advancements in artificial intelligence (AI), particularly large language models (LLMs). We demonstrate how AI can automate and enhance cyberattacks to identify the most pressing dangers and present feasible mitigation strategies. The study is divided into two main branches: attacks targeting hardware and software systems and attacks focusing on system users, such as phishing. The first paper of the thesis identifies research communities within computer security red teaming. We created a Python tool to scrape and analyze 23,459 articles from Scopus's database, highlighting popular communities such as smart grids and attack graphs and providing a comprehensive overview of prominent authors, institutions, communities, and sub-communities. The second paper conducts red teaming assessments of connected devices commonly found in modern households, such as connected vacuum cleaners and door locks. Our experiments demonstrate how easily attackers can exploit different devices and emphasize the need for improved security measures and public awareness. The third paper explores the use of LLMs to generate phishing emails. The findings demonstrate that while human experts still outperform LLMs, a hybrid approach combining human expertise and AI significantly reduces the cost and time requirements to launch phishing attacks while maintaining high success rates. We further analyze the economic aspects of AI-enhanced phishing to show how LLMs affect the attacker's incentive for various phishing use cases. The fourth study evaluates LLMs' potential to automate and enhance cyberattacks on hardware and software systems. We create a framework for evaluating the capability of LLMs to conduct attacks on hardware and software and evaluate the framework by conducting 31 AI-automated cyberattacks on devices from connected households. The results indicate that while LLMs can reduce attack costs, they do not significantly increase the attacks' damage or scalability. We expect this to change with future LLM versions, but the findings present an opportunity for proactive measures to develop benchmarks and defensive tools to control the misuse of LLMs.

Abstract [sv]

Moderna cyberattacker förändras snabbt som följd av framsteg inom artificiell intelligent (AI), särskilt via stora språkmodeller (LLM:er). Vi demonstrerar hur AI kan automatisera och förbättra cyberattacker för att identifiera de största hoten och presenterar strategier för att motverka dem. Studien är uppdelad i två delar: attacker riktade mot hårdvaru- och mjukvarusystem samt attacker fokuserade på systemanvändare, likt phishing. Avhandlingens första artikel identifierar forskningsgrupper inom red teaming. Vi skapade ett Python-verktyg för att hämta och analysera 23,459 artiklar från Scopus databas, vilket gav en översikt av framstående författare, institutioner och utvecklingen av olika grupper och sub-grupper inom forskningsområdet. Avhandlingens andra artikel genomför red teaming-tester av uppkopplade enheter från moderna hushåll, exempelvis uppkopplade dammsugare och dörrlås. Våra experiment visar hur lätt angripare kan hitta sårbarheter i enheter och betonar behovet av förbättrade säkerhetsåtgärder och ökad allmän medvetenhet. Den tredje artikeln utforskar användningen av LLMs för att generera phishing-meddelanden. Resultaten visar att mänskliga experter fortfarande presterar bättre än LLMs, men en hybridmetod som kombinerar mänsklig expertis och AI reducerar kostnaderna och tiden som krävs för att lansera nätfiskeattacker och behåller hög kvalitet i meddelandena. Den fjärde studien utvärderar LLM:ers potential att automatisera och förbättra cyberattacker på hårdvaru- och mjukvarusystem. Vi skapar ett ramverk för att utvärdera LLM:ers förmåga att genomföra attacker mot hårdvara och mjukvara och utvärderar ramverket genom att genomföra 31 AI-automatiserade cyberattacker på enheter från uppkopplade hushåll. Resultaten indikerar att LLM:er kan minska attackkostnaderna, men de medför inte en märkvärd ökning av attackernas skada eller skalbarhet. Vi förväntar oss att detta kommer att förändras med framtida LLM-versioner, men resultaten presenterar en möjlighet för proaktiva åtgärder för att utveckla riktmärken och försvarsverktyg för att kontrollera skadlig användning av LLMs.

Place, publisher, year, edition, pages
Stockholm: KTH Royal Institute of Technology, 2024. p. x, 71
Series
TRITA-EECS-AVL ; 2024:68
Keywords
Computer security, Red teaming, phishing, artificial intelligence, large language models
National Category
Computer Sciences
Research subject
Computer Science
Identifiers
urn:nbn:se:kth:diva-353243 (URN)9789181060409 (ISBN)
Public defence
2024-10-10, https://kth-se.zoom.us/j/61272075034, D31, Lindstedtsvägen 9, Stockholm, 13:00 (English)
Opponent
Supervisors
Note

QC 20241004

Available from: 2024-09-19 Created: 2024-09-18 Last updated: 2024-10-21Bibliographically approved

Open Access in DiVA

No full text in DiVA

Other links

Publisher's full textScopus

Authority records

Heiding, Fredrik

Search in DiVA

By author/editor
Heiding, Fredrik
By organisation
Network and Systems Engineering
In the same journal
IEEE Access
Computer Sciences

Search outside of DiVA

GoogleGoogle Scholar

doi
urn-nbn

Altmetric score

doi
urn-nbn
Total: 315 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf