딥페이크와 증거재판주의 - 딥페이크 생성·탐지와 디지털 증거에 관한 이해를 바탕으로

Deep generative artificial intelligence models produce countless deepfake videos, images, and audio, which are indistinguishable from reality. This technological development can undermine the probative force of video evidence. Once defense is made that evidence has been forged using deepfake, assessing its admissibility and probative value becomes difficult and proving the authenticity of video evidence may consume excessive time and resources. Since the introduction of the Generative Adversarial Network in 2014, generation models have advanced significantly. Early face-swap systems required training a separate model for each pair of individuals, but now one can perform a face-swap with minimal data, even a single photograph pair. High-quality face replacement now so closely mimics reality that its detection is extremely challenging with bare human eyes, and modern tools such as Sora can generate videos purely from text prompts. Open-source and commercial deepfake applications are now readily available to the general public. Yet deepfake detectors have not delivered strong performance in (i) transferability(applying to unseen data), (ii) interpretability(explaining why they reached a given conclusion), and (iii) robustness(maintaining consistent detection under compression, post‑processing, or other real-world distortions). Humans achieve around 70% accuracy in detecting deepfakes, yet they tend to overestimate their ability to identify deepfakes. This bias poses a grave risk of mistakenly convicting defendants in judicial proceedings based on misidentified deepfakes. In those cases, U.S. courts held that simply raising the possibility of deepfake manipulation without concrete proof does not preclude admissibility of the alleged deepfake evidence. U.S. courts have already examined some deepfake defenses. In those cases, U.S. courts held that where there is “mere speculation that contamination or tampering could have occurred, it is not an abuse of discretion to admit the evidence.” Amendments to the Federal Rules of Evidence have been proposed, including requiring a judge’s gatekeeping role, placing the burden of proof on the challenger before admitting potentially deepfake-generated evidence, and requiring that the probative force of evidence must outweigh its potential prejudicial impact. This is to strike a balance between the prevention of deceptive defenses aimed at excluding valid evidence, and the protection of jurors from prejudicial deepfake material. We can expect similar deepfake defenses on trial. The traditional tests of identity, chain of custody and reliability do not guarantee the validity of evidence. Deepfake detection models often rely on ‘blackbox’ deep learning architectures, making them difficult to accept as scientific evidence in the strict sense. Courts may order either pretrial hearings or expert examination procedures for contested video evidence. While expert evaluations sometimes uncover deepfake manipulation, in other cases judges must assess authenticity based on corroborating evidence, rather than relying solely on technical or scientific method. Judges must not decide on the video’s authenticity based solely on their impression. When video evidence can be viewed as a statement, the court should first adjudicate deepfake defense. If the grounds for the deepfake defense are prima facie persuasive, the court will have to consider ruling out the participatory trial to mitigate the risk of prejudgment. Adjudicating deepfake defense is a dynamic assessment of deepfake and its detection technology, as of the time of submission. While courts face new challenges with the advent of deepfake, they can uphold the principle of evidence while pursuing the substantive truth.