Perceptual and computational detection of face morphing.

Area of Science:

• Forensic science
• Computer vision
• Human-computer interaction

Background:

• Identity documents increasingly use digital images.
• Facial morphing technology blends two individuals' images, creating a new identity.
• This technique poses a novel threat for identity fraud and document security.

Purpose of the Study:

• To assess human and machine-learning-based facial recognition system's ability to detect digitally morphed facial images.
• To evaluate the effectiveness of current identification verification methods against sophisticated facial morphing techniques.

Main Methods:

• Generation of a diverse dataset of high-quality digital facial morphs from passport-style photographs.
• Human participant testing for detecting morphed images (same vs. different individuals, identifying morphed faces).
• Evaluation of modern machine learning-based facial recognition algorithms on the morphed image dataset.

Main Results:

• Human participants demonstrated significant difficulty in accurately detecting facial morphing.
• Participants struggled to determine if side-by-side faces belonged to the same individual.
• Facial recognition systems also exhibited limitations in distinguishing between original and morphed facial images.
• Both human and AI methods showed low accuracy in identifying morphed facial images.

Conclusions:

• Current human and automated systems are vulnerable to identity theft using morphed facial images.
• The study highlights a critical gap in digital identity verification security.
• A promising computational technique for detecting morphed images was identified, offering potential future solutions.