Learned Image Compression for Deepfake Detection

Le résumé fourni par la source

Learned image compression introduces new challenges for image forensics, as deepfake detectors may mistakenly classify authentic images as artificially generated due to compression artifacts or misidentify synthetic images that have undergone neural compression as real. In this paper, the first neural image compression system specifically designed for deepfake detection is presented. To achieve this, spatial-frequency modulation adapters are integrated into an existing image compression architecture, eliminating the need to retrain the underlying codec. Images decoded with the proposed method achieve detection performance comparable to uncompressed images while also exhibiting superior perceptual quality than images reconstructed from conventional image compression.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.