A new method for identifying deepfake videos achieves 97.8% accuracy while consuming less energy than digital neural networks with comparable performance. The technique relies on the physics of light, rather than conventional digital processing, to distinguish fake audiovisual content from authentic material.

Deepfakes, artificially generated or manipulated videos that can make people appear to say or do things that never happened, have become a growing concern for content verification and information security. Deep neural network detection systems have proven effective, but they typically demand considerable energy consumption and substantial computing resources.

The alternative presented uses the physics of light to carry out the classification task at a lower energy cost. According to the researchers, the system achieves 97.8% accuracy in detecting fake videos, a result comparable to digital neural networks with similar performance, but with lower energy consumption.

The combination of high accuracy and energy efficiency opens the door to practical applications in resource-limited devices, where running large digital models is unfeasible. The work suggests that approaches inspired by physical phenomena, rather than purely digital computation, can compete in machine-learning tasks relevant to society.