Abstract:

Audio deepfakes are artificial speech recordings that use machine learning algorithms to seem like actual human voices. These technologies represent major threats such as disinformation and fraud, making deepfake detection a critical skill for computing students. However, teaching audio deepfake detection is challenging, especially as the quality of fake audio improves. Traditional lecture-based methods are not effective at engaging students with diverse, complex, and interdisciplinary content. Scenario-based learning using a caselet-based approach presents an effective pedagogical method for the education of deepfake detection. Caselets provide structured, real-world scenarios that integrate technical concepts with practical applications, allowing students to develop both analytical skills and security awareness through authentic problem-solving experiences. However, limited research exists on the successful implementation of caselet-based approaches in influencing learning outcomes in cybersecurity education.

We studied whether Caselets - structured, scenario-based learning exercises - can effectively teach these concepts by drawing upon knowledge from both computer science and linguistics. We tested our approach with 53 undergraduate students in two computing courses using different teaching methods. Course A (n=29) used guided instruction with linguistic feature training, while Course B (n = 24) used collaborative interactive learning. The Caselet "Stopping Scam Calls with Deepfake Detection" was used.

This Caselet presents a banking fraud scenario where students develop automated detection systems in three phases:
(1) problem formulation and model selection,
(2) feature engineering using linguistic features and preprocessing, and
(3) model validation including overfitting analysis and confusion matrix interpretation.

Students completed pre- and post-class surveys measuring their self-reported knowledge on 5-point Likert scales, plus three open-ended questions about new learning, lesson improvements, and topics for further exploration.

The sample consisted mainly of undergraduates ages 18-24 (94.5%) with somewhat balanced gender representation (60% male, 36.4% female). The participants came from a variety of academic majors and they had varying levels of cybersecurity backgrounds. The Shapiro-Wilk test showed that the data were not normally distributed, thus we had to use non-parametric statistical approaches. Wilcoxon signed-rank tests were applied to assess if knowledge gains were achieved from pre-test to post-test.

The Caselet approach employed in this study led to significant knowledge gains for undergraduate students in the area of audio deepfake detection, while helping them understand concepts from machine learning and linguistics. Students called for more practice, real-world links, interaction, linguistic focus, and research ties. Caselets effectively supported interdisciplinary learning. The approach shows promise for scaling to other computing courses that involve emerging technologies with societal implications.

Thematic analysis revealed three key recommendations for caselet design:
(1) provide more audio sample libraries for diverse practice opportunities,
(2) integrate real-world security applications to build institutional awareness, and
(3) incorporate gamification elements.

These findings provide actionable insights for educators implementing scenario-based cybersecurity education in computing curricula.

Keywords:

AI deepfake caselet study, Cybersecurity education.