Evaluating the real-world robustness of face-swap detection models under compression and noise

Recent advances in generative adversarial networks (GANs) and autoencoding techniques have significantly improved the realism of face-swap and deepfake media, creating substantial challenges for digital media authentication. Although existing deepfake detection models achieve high accuracy on benchmark datasets, their robustness under real-world media degradations remains insufficiently explored.

This study systematically evaluates the resilience of five leading face-swap detection models-XceptionNet, MesoNet, FSD-GAN, FakeTracer, and a Hybrid + Landmark approach-under four common distortions: JPEG compression (quality levels 20-90), Gaussian noise (σ = 0.01-0.05), motion blur (kernel size 3-15), and video encoding artefacts (bitrate 50-500 kbps). Experiments were conducted using the FaceForensics++ dataset (1,000 videos: 720 training, 140 validation, and 140 testing) and Celeb-DF v2 (590 videos: 400 training, 90 validation, and 100 testing). Performance was assessed using accuracy, F1-score, area under the curve (AUC), and degradation rate (Δ) between clean and distorted conditions.

The results demonstrate a substantial reduction in detection performance under degraded conditions. Average accuracy declined from 94.7% on clean data to 67.8% on distorted data, corresponding to an overall degradation rate of -26.9%. JPEG compression and motion blur caused the most significant performance drops, with reductions of up to 35%, particularly for lightweight CNN-based detectors. In contrast, FSD-GAN and FakeTracer exhibited greater robustness, maintaining degradation rates of no more than -15% due to their latent fingerprinting and trace embedding mechanisms.

The findings highlight the limitations of current deepfake detection systems when deployed in real-world environments where media distortions are prevalent. The study emphasizes the need for distortion-aware training strategies, cross-condition benchmarking, and deployment-oriented evaluation protocols. Furthermore, a dual-branch framework integrating a Vision Transformer (ViT) for spatial artefact detection with a Recurrent Neural Network (RNN) or Temporal Convolutional Network (TCN) for temporal coherence modelling is proposed as a promising direction for improving the robustness and reliability of future deepfake detection systems.