Detection of AI-Generated Social Engineering in Code-Switched Darija-French Communications
Detection of AI-Generated Social Engineering in Code-Switched Darija-French Communications
Description
Details
Context and Problem Statement
Phishing detectors often rely on linguistic cues learned from English corpora. Generative AI removes many of these cues, while Moroccan communication frequently mixes Darija, French, and Modern Standard Arabic, a pattern rarely represented in training data.
Research Question
Which linguistic and structural signals remain useful for distinguishing AI-generated social-engineering messages from legitimate communication under code-switching, and do they remain robust under adversarial paraphrasing?
Proposed Approach
Construct a balanced corpus of anonymized legitimate messages, real phishing samples, and generated variants annotated for manipulation strategies. Compare multilingual encoders, stylometry, and structural features such as headers and link reputation, followed by adversarial paraphrase testing.
Expected Contribution
A North-African multilingual social-engineering benchmark and a robustness analysis under AI-assisted paraphrasing.
Expected Prototype
A message-scoring service suitable for integration into email or messaging gateways, with explanation and adjustable risk thresholds.
Datasets
Nazario, Enron, and a consent-based collected multilingual corpus.
Challenges
Privacy protection, imperfect labels, potential bias against legitimate language varieties, and the evolving generation-detection arms race.
Research Question
Innovation
Expected Deliverable
Technologies
Required Skills
- Multilingual Natural Language Processing
- Email and Messaging Security
- Data Collection and Anonymization Methodology
- Adversarial Model Evaluation
Datasets
- Nazario Phishing Corpus
- Enron Email Dataset
- Collected and anonymized multilingual corpus