Adaptive Gated Deepfake Detection for Low-Resolution and Resource-Constrained Environments

With the rapid advancement of generative models, deepfake technology has made manipulated facial media increasingly realistic and difficult to distinguish from authentic content. Consequently, deepfake detection has become a major issue for digital forensics and media verification. Benchmark datasets like FaceForensics++ and Celeb-DF have supported standardized evaluation in this area. Despite progress, detecting deepfakes under low-resolution and resource-constrained conditions remains challenging.

Deepfake detection models often rely on high-quality inputs, fixed inference paths, and computationally expensive architectures, limiting their use in low-resolution and resource-constrained settings. In real-world scenarios, images and video frames are often compressed, resized, blurred, or captured at inconsistent resolutions, which can reduce or distort fine-grained facial artifacts useful for detecting manipulation.

AdaGate-DF combines RGB and frequency-domain features with an adaptive inference mechanism, significantly improving detection performance on low-quality inputs. By routing based on quality features, it allows high-quality images to exit early, saving computational resources. This approach is innovative compared to existing methods, especially in resource-limited environments.

• �� AdaGate-DF employs a dual-branch architecture combining RGB and DCT features.
• �� Utilizes a multi-exit inference mechanism that adapts the inference path based on image quality cues.
• �� Routes samples based on quality features, reducing computational resource usage.
• �� Evaluated on Celeb-DF and FF++ datasets, comparing performance with existing methods.

Experiments were conducted on Celeb-DF and FF++ benchmark datasets to evaluate performance under different input resolutions. Baselines include MaD-CoRN, DefakeHop++, and ShuffleNetV2. Key evaluation metrics include AUC, accuracy, precision, recall, and F1-score.

On the Celeb-DF dataset, AdaGate-DF achieves an AUC of 0.9370, outperforming MaD-CoRN and DefakeHop++ while maintaining low inference latency. At 384×384 resolution, the AUC reaches 0.9708, demonstrating performance improvement with increased resolution. On the FF++ dataset, AdaGate-DF remains effective under class imbalance, showing competitive results with evaluated models.

AdaGate-DF is applicable in scenarios such as digital forensics, media verification, and maintaining online trust, especially in resource-constrained environments. Its adaptive inference mechanism makes it more efficient in processing large volumes of images and video frames.

In the imbalanced FF++ dataset, despite good AUC performance, there may still be a risk of bias towards the majority class. Performance may degrade in extremely low-resolution scenarios. Future research can explore generalization capabilities across more datasets and performance optimization in real-time applications.