A IA consegue detetar vídeos deepfake ao analisar inconsistências microscópicas nos padrões de piscar ?

Os investigadores de IA descobriram que os vídeos sintéticos mostram, de forma consistente, dinâmicas de piscar de olhos não naturais. Estes sistemas utilizam análise de vídeo de alta resolução para identificar inconsistências invisíveis ao olho humano. A técnica funciona na maioria dos métodos atuais de geração de deepfakes. No entanto, novos ataques adversariais já estão a ser desenvolvidos para contornar essa deteção.

Background

Current deepfake detection methods do analyze subtle physiological cues, and blinking patterns have been explored because synthesized faces often produce unnaturally consistent or infrequent blinks. Research shows that deep neural networks can learn to detect these microscopic inconsistencies by examining blink frequency, duration, and eyelid motion dynamics, sometimes achieving high accuracy on controlled datasets (Li, Y., et al. "Exposing AI-Generated Faces by Detecting Eye Blinking Anomalies." 2022 IEEE International Conference on Multimedia and Expo (ICME)). However, as generative models improve, attackers can refine blinking behavior to evade such detectors, making this approach increasingly unreliable as a standalone defense. Performance varies widely across lighting conditions, head poses, and video compression, limiting real-world applicability. New adversarial attacks are already being developed to bypass such detection.