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

Advances in synthetic media, including generative adversarial networks (GANs), diffusion models, and face manipulation tools, have led to a growing need for robust deepfake detection systems. However, most single deepfake detectors struggle to maintain robustness as the methods used for synthesis change. To address this, the researchers introduce DeepFakeBuster, a confidence-calibrated adaptive ensemble framework designed for deepfake image detection.

This framework combines various deep learning models that detect different forensic cues such as spatial inconsistencies, boundary artifacts, noise residuals, semantic consistency, and frequency-domain features. Unlike traditional ensemble methods that use static averaging, DeepFakeBuster uses a reliability-aware adaptive fusion approach.

This dynamic method adjusts the contribution of each detector based on reliability priors derived from validation and input-specific confidence estimates.

In extensive testing on a dataset containing 192,000 authentic and manipulated images, the ensemble model significantly surpassed both individual detectors and static fusion baselines. It achieved an overall accuracy of 97.8% under the tested conditions. Moreover, the framework includes an interpretable forensic analysis module that generates visual and quantitative indicators highlighting manipulation-sensitive areas.

The study highlights confidence-aware heterogeneous ensemble learning as a promising approach for robust deepfake detection. The authors express gratitude to their institutions for research facilities and computational support, as well as to the providers of public datasets and the open-source community for tools and frameworks that aided their research.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.