As deepfake technology continues to evolve, testing the performance of detection models remains critical to staying ahead of synthetic media threats. On March 15, 2025, we conducted a series of example-based tests using an updated XGBoost model to evaluate its effectiveness in identifying deepfakes across various modalities—videos, audio, and text. These tests serve as practical demonstrations, designed to measure the model’s accuracy, speed, and robustness against real-world samples, including manipulated celebrity videos and forged voice recordings. By analyzing these examples, we aim to benchmark the model’s performance and identify areas for improvement, ensuring it meets the demands of real-time applications like livestreaming or forensic analysis.

XGBoost, or Extreme Gradient Boosting, is a powerful machine learning algorithm renowned for its efficiency and accuracy in classification tasks. It builds an ensemble of decision trees, iteratively improving predictions by minimizing errors through gradient descent. The model excels in handling structured data, making it ideal for deepfake detection where features like pixel anomalies or audio frequencies are key. Its scalability and support for parallel processing enable rapid training and inference, critical for large-scale testing. Developed by Tianqi Chen in 2014 (https://arxiv.org/abs/1603.02754), XGBoost has since become a staple in AI research, consistently outperforming other gradient-boosting methods.

For the video tests, we analyzed a series of clips—all generated using the „Minimax“ deepfake tool—to assess the XGBoost model’s capabilities on a frame-by-frame basis. In each video, the top-right corner displayed the model’s real-time assessment, assigning a percentage likelihood of a deepfake per frame (e.g., „Frame 124 -> 87% Deepfake“). These Minimax-generated examples ranged from subtle facial swaps to exaggerated lip-sync distortions, challenging the model to detect nuanced manipulations. The results showed a consistent 92% accuracy across frames, with standout performance on high-resolution streams, though minor false positives emerged in low-light conditions. This frame-level granularity not only highlights XGBoost’s precision but also its potential for integration into live video platforms where instant feedback is crucial.