Deepfakebuster: a confidence-calibrated adaptive ensemble framework for robust Deepfake image detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Generative adversarial networks (GANs) and diffusion models create realistic synthetic media.
• Existing deepfake detectors lack robustness due to reliance on specific forensic cues and poor adaptation to evolving synthesis techniques.

Purpose of the Study:

• To develop a robust and adaptive deepfake detection system.
• To improve the accuracy and reliability of deepfake detection by overcoming limitations of single-model approaches.

Main Methods:

• Developed DeepFakeBuster, a confidence-calibrated adaptive ensemble for deepfake image detection.
• Fused heterogeneous deep learning models detecting complementary forensic cues (spatial, boundary, noise, semantic, frequency-domain).
• Utilized reliability-aware adaptive fusion, dynamically adjusting detector contributions based on confidence estimates.

Main Results:

• Achieved an overall accuracy of 97.8% on a dataset of 192,000 images.
• Significantly outperformed individual detectors and static fusion baselines.
• Included an interpretable module for visual and quantitative analysis of manipulation-sensitive regions.

Conclusions:

• Confidence-aware heterogeneous ensemble learning is a promising direction for robust deepfake detection.
• DeepFakeBuster demonstrates superior performance and adaptability compared to existing methods.
• The interpretable module aids in understanding and validating detection results.