Low Resolution Facial Manipulation Detection

Abstract

Detecting manipulated images and videos is an important aspect of digital media forensics. Due to severe discriminative information loss caused by resolution degradation, the performance of most existing methods is significantly reduced on low resolution manipulated images. To address this issue, we propose an Artifacts-Focus Super-Resolution (AFSR) module and a Two-stream Feature Extractor (TFE). The AFSR recovers facial cues and manipulation artifact details using an autoencoder learned with an artifacts focus training loss. The TFE adopts a two-stream feature extractor with key points-based fusion pooling to learn discriminative facial representations. These two complementary modules are jointly trained to recover and capture distinctive manipulation artifacts in low resolution images. Extensive experiments on two benchmarks including FaceForensics++ and DeepfakeTIMIT, evidence the favorable performance of our method against other state-of-the-art methods.