Scoop: Mitigation of Recapture Attacks on Provenance-Based Media Authentication

Continuous advances in photo and video manipulation yield increasingly sophisticated deepfakes that greatly endanger societal perception of reality. Deepfake detection is an intuitive and natural research direction, which is unfortunately shaping up to be a never-ending arms race. An alternative promising direction is provenance assertion, which blends hardware-based secure camera design with the cryptographic means of authenticating the source of visual content and any post-processing (e.g., filters) applied to it.

This work starts by highlighting a very effective attack type, called a recapture attack, against all provenance-based techniques. In such an attack, the adversary displays fake content on some form of a screen (e.g., TV, projector, or computer screen) or surface (e.g., cardboard, canvas, or paper) and uses a provenance-asserting secure camera device to capture photos and videos of the displayed content.

We then introduce Scoop, a systematic solution for mitigating recapture attacks. Scoop leverages state-of-the-art depth sensing technologies as well as learning-based depth estimation to detect misleading recaptures, i.e., a recaptured photo or video where the presence of a display medium is not visually identifiable.

We implement Scoop on both iOS and Android platforms (Apple iPhone 14 Pro and Samsung Galaxy S20 Plus), using their built-in depth sensors. To evaluate the effectiveness of Scoop, we construct a first-of-its-kind dataset consisting of 78 recapture attack scenarios. Our results show that Scoop achieves as high as ≈ 95% accuracy on the iPhone and 74% accuracy on the Samsung phone.