> DISSERTATION: HONOURS_PROJECT_2025

Linguistic
Realism.

Evaluating brand imitation in AI-generated phishing. A 23,000-word investigation into how Large Language Models (LLMs) automate the erosion of digital trust.

Recommended for Publishing Supervisor Endorsed

01. The Problem Space

From Sony to ChatGPT.

My literature review traced the evolution of phishing from the manual, high-effort spear-phishing of the Sony 2014 Breach (costing $100m) to today's automated AI arms race.

The core issue is scalability. What once required a team of hackers to mimic a brand's voice can now be done in seconds by a generative model.

The V-Triad Framework

I applied this psychological model to explain why we click. Attackers exploit Credibility (brand logos) and Urgency (deadlines) to bypass logical filters.

Historical_Context

  • 2014 Sony Breach: Manual, high-effort, high-impact.
  • 2020 GPT-3 Release: The beginning of automated text generation.
  • 2024 Current State: 1.3 million phishing attacks per quarter.

02. Methodology Lab

Mathematical Forensics.

How do you measure "realism"? I didn't just read emails; I converted them into data points to measure their similarity mathematically.

BERT Semantic Embedding

Think of BERT as a translator that turns sentences into GPS coordinates. I converted 60+ emails into numerical vectors using all-MiniLM-L6-v2.

By calculating the distance between the "Real Brand" vector and the "AI Clone" vector, I proved they occupy almost the exact same semantic space.

Calculated_Cosine_Similarity: 0.8490

Empath Tone Analysis

I used Empath to analyze lexical categories like 'Money', 'Urgency', and 'Trust'.

The findings showed that AI is excellent at mimicking corporate authority but often lacks the subtle "entropy" or randomness of human writing—a key marker for future detection tools.

03. The Verdict

49.2%

The Misclassification Trap.

The most alarming finding: Nearly half of study participants mistook AI-generated branded phishing for legitimate communication.

This 50/50 split proves that human intuition is no longer a reliable firewall. When the "Linguistic Fingerprint" is cloned perfectly, the user cannot defend themselves.

The Confidence Gap (Fig 11)

High Confidence (Wrong) Low Confidence

Participants were often "Very Confident" that a phishing email was real.

"We are entering an era where detection must rely on technical infrastructure verification, as the visual and linguistic cues we rely on have been fully compromised."