Genuine and Fake Facial Emotion Dataset (GFFD-2025)

The Dual-Task Emotion–Authenticity Facial Expression Dataset (GFFD-2025) is a carefully curated collection of facial images created to support research in emotion recognition and authenticity detection. Unlike traditional emotion datasets, it focuses not only on identifying which emotion a person expresses but also on whether the expression is genuine or acted, contributing to studies in artificial intelligence, affective computing, and human–computer interaction.

A total of 2,224 raw facial images were initially collected from voluntary participants. After quality assessment and manual verification, a subset was refined and curated for further research. The dataset repository includes approximately 1,900 raw facial images and around 1,500 cropped and augmented images, representing the cleaned and extended version of the original collection.

The dataset covers seven primary emotions: Angry, Disgust, Fear, Happy, Neutral, Sad, and Surprise; each subdivided into two authenticity categories: Genuine and Fake (Acted). Images were captured under controlled indoor conditions to ensure consistent lighting, neutral backgrounds, and stable face positioning. Genuine expressions were elicited via emotional recall or audiovisual stimuli, while fake expressions were intentionally acted. All data collection sessions were supervised by a certified psychologist to ensure ethical compliance and emotional validity.

Images were reviewed and labeled following micro-expression research principles, considering subtle cues such as eye involvement, facial symmetry, muscle tension, and temporal dynamics to distinguish genuine from acted expressions. Curated images were standardized to 224×224 pixels for compatibility with common deep learning frameworks.

To enhance dataset diversity and model robustness, images underwent preprocessing and augmentation, including rotation (±30°), width and height shifts (0.2), shear (0.15), zoom (0.2), horizontal flipping, random brightness and contrast adjustments, and normalization to the [0,1] range.

This dataset offers a practical benchmark for research in emotion recognition, authenticity detection, human behavior analysis, multitask learning, and explainable AI, enabling development of models sensitive to subtle psychological authenticity cues.

Data collection and labeling were conducted at Daffodil International University, Dhaka, Bangladesh, under strict ethical guidelines with informed consent from all participants. Sessions were supervised to ensure participant comfort and authenticity.

Supervisor:
Md. Mizanur Rahman
Lecturer, Department of Computer Science and Engineering
Daffodil International University, Dhaka, Bangladesh
Email: [e-mail]

Data Collectors:
Sarah Tasnim Diya (Email: [e-mail])
Most. Jannatul Ferdos (Email: [e-mail])
Department of Computer Science and Engineering, Daffodil International University, Dhaka, Bangladesh.