This is for a project I'm working on: an open-source CAPTCHA. A prover receives a random challenge and must submit a video of themselves performing specified actions. These might involve moving their hands in a particular sequence, manipulating objects, or performing several simultaneous actions. The submitted video is evaluated by a computer-vision pipeline consisting of multiple verifiers. Together, they produce a binary pass/fail result. Importantly, the verifiers are entirely white-box: their architectures, weights, source code and acceptance thresholds are public. They are also deterministic. The checks need not be simple deepfake detectors. They could verify geometric properties of objects, contact between objects and the performer's hands, precise finger movements, physical plausibility, deformation, multi-view consistency, etc. My question is: Can such a verifier pipeline force an attacker to incur substantial computational costs for every successfully generated video, or can an attacker eventually train a generator to satisfy the public checks with very little additional inference cost? I see two competing considerations: An attacker can inspect the public verifiers, identify weaknesses and optimize a generator against them. Verifier feedback could potentially be distilled into the generator's weights, avoiding the need to run the verifiers during every generation. On the other hand, satisfying a large conjunction of sophisticated geometric, physical and temporal co…

Full article content could not be extracted automatically. Read the original below.